Other Leading Models: Qwen, Kimi, GLM, Llama, Mistral and More

Key takeaways

Facts are as of 12 October 2026 and come from the sources reviewed for this research. Licence texts, prices, index scores and download counts were re-checked against the primary pages on 11 October 2026, and vendor price pages, Xiaomi's repositories, cloud-region tables and Artificial Analysis's leaderboard against primary sources on 12 October 2026. Vendor claims and single-source items are labelled as such. INR figures use ₹96 per USD. The licence summaries are a planning-level reading of licence texts, not legal advice. Claude, GPT, Gemini, Grok and DeepSeek each have their own document (Claude, OpenAI, Gemini, Grok, DeepSeek). The cross-vendor ranking is in Top models, and the wider context is in the research index, Landscape, Timeline, Future outlook and the Glossary.

• Chinese labs now hold most of the open-weight frontier. The largest downloadable models are Moonshot's Kimi K3 (2.78 trillion parameters) and Alibaba's Qwen3.8-2.4T-A95B. On Artificial Analysis's own leaderboard (Intelligence Index v4.3.2, live 12 Oct 2026), the best open-weight scores are Xiaomi's MiMo-V2.6-Pro (46.3), GLM-5.3 (44.8) and Kimi K3 (43.6). MiMo is tagged MIT in its Hugging Face model-card metadata, but its public repos contain no LICENSE file. The top closed model, Claude Opus 5.5 (Max), scores 57.6.
• Traffic follows price. Six of the ten models that processed the most OpenRouter tokens in the latest week came from Chinese companies: two DeepSeek models, GLM 5.3 Flash, Xiaomi MiMo-V2.6-Flash, Tencent Hy4-preview and StepFun Step 5 Preview. Most of them are low-price tiers.
• Licences fall into three groups. Apache 2.0 or MIT with no thresholds (Qwen3.8-27B, Qwen3.5, Tencent Hy3 and Hy4-preview, StepFun Step 3.x Flash, GLM-5/5.2/5.3-Flash, Xiaomi MiMo-V2.6 (MIT only in model-card metadata: no LICENSE file in the public Pro-RL and Flash-RL repos, and the plain Pro and Flash repos are gated, HTTP 401; treat its terms as unconfirmed), gpt-oss, Granite, Phi). Custom licences with revenue or user triggers (Qwen3.8-Max at US$50M of group revenue for firms running hosted-model or AI work-assistant businesses, Kimi K3 at US$20M of group revenue for hosted-model businesses, Llama 4 at 700M monthly users, Liquid LFM with a US$10M revenue cap). And MiniMax M3, whose base grant is non-commercial: any commercial use needs a "Built with MiniMax M3" label and a notice to MiniMax, or written authorisation once the products using it earn US$20M a year.
• Western open weights compete on provenance and small sizes. Meta's top model is now closed (Muse Spark), Mistral promises Large 4 weights (1T total, 52B active) by end-October, NVIDIA Nemotron 3 and IBM Granite 4.2 are the enterprise-friendly families, and gpt-oss-120b remains the cheap hosted reference at $0.037 / $0.17 per 1M tokens.
• Hosted open models cost almost nothing per task. Under the costs analysis's assumed invoice-extraction profile, 1,000 documents cost about ₹25 on gpt-oss-120b and ₹89 on GLM-5.3-Flash. The same batch costs about ₹2,160 on Kimi K3 at Moonshot's list price.
• Hardware sets the practical ceiling for self-hosting. A 128 GB desktop box (DGX Spark, about ₹4.5 lakh at its US list price) runs gpt-oss-120b at about 59 tokens per second for a single user. Models of about 1T parameters need a 748 GB-class workstation even at 4-bit, and the 2.4T–2.8T open models (Qwen3.8-2.4T, Kimi K3) need clusters, so for those a hosted API is the realistic route.
• The main risk with Chinese models is where the data goes. The cheapest endpoints are hosted in the PRC. Evidence of censorship in the sources reviewed is thin: one AFP test of Ernie Bot. Self-hosting Apache or MIT weights in India or in your own region removes the data-flow risk. AWS Bedrock also routes Kimi K3 through an India-only profile (Mumbai and Hyderabad). It runs older open models In-Region in Mumbai: Qwen3, Kimi K2.5, GLM 4.7/5, MiniMax M2.x, Mistral Large 3, Gemma 3, Llama 3 and gpt-oss. The newest releases (Qwen3.5/3.8, GLM-5.3, MiniMax M3, MiMo) are not on AWS's Mumbai list as read on 12 Oct 2026. India's DPDP core duties apply from 13 May 2027.

The master table: 30 model lines at a glance

The table covers the families that matter for automation outside the five big vendors, with OpenAI's gpt-oss and Google's Gemma as open-weight reference points. Prices are USD per 1M tokens (input / output). "OR" marks an OpenRouter host price, which is an aggregator price, not the vendor's list price. "n/s" means not stated in the sources reviewed, "n/d" not disclosed, and "from name" means read from the model ID.

Sources for the table: Hugging Face API and licence files (Qwen org, Qwen3.8-Max licence, Kimi K3 licence, MiniMax M3 licence, zai-org, Tencent Hy3, Tencent Hy4-preview, Step-3.7-Flash, XiaomiMiMo, MiMo-V2.6-Pro-RL, Llama 4 licence, LFM licence, gpt-oss card, Phi-4 card, Phi-4-mini card); vendor price pages (Alibaba Model Studio, Kimi platform, Z.ai, MiniMax, Xiaomi MiMo, AWS Bedrock price list); the OpenRouter models API and rankings; Mistral Large 4; IBM Granite.

Three price bands stand out. Hosted small and mid-size open models (gpt-oss, Nemotron Nano and Super, Granite, Phi-4, Gemma 4, Hy3) cost under $0.10 per 1M input tokens. Chinese mid-tier and Western workhorse models cost $0.15 to $1.50. The largest Chinese and Meta models cost $1.25 to $3, still well below the $10 that the most capable closed models charge (OpenRouter models API). A 1M-token context no longer sets a model apart, so the choice comes down to licence, hosting location and measured task quality.

Chinese makers lead on breadth, scale and price

Chinese labs set the pace in 2026 in two ways. They release often, and they publish weights at sizes no Western lab matches. Zhipu (HKEX: 2513) and MiniMax (HKEX: 0100) listed in Hong Kong on 8 and 9 January 2026 (Madrimasd) and joined the Hang Seng Tech Index on 5 June (AIWeekly). Compute is the constraint. The US approved H200 exports to China in January 2026, and China blocked the shipments within days. Earlier, on 17 September 2025, the Cyberspace Administration of China had told companies including ByteDance and Alibaba not to buy NVIDIA's RTX Pro 6000D (Wikipedia, Nvidia). The labs' mixture-of-experts designs fit that scarcity: each token activates only a small share of the weights, such as 17B of Qwen3.5's 397B or 21B of Hy3's 295B. No source reviewed states that link, so treat it as an inference. Regional context is in Asia-Pacific and Hardware.

Alibaba Qwen ships every size, with one custom licence at the top

Qwen is the broadest family. Qwen3.5 arrived in mid-February 2026 with open weights at 397B-A17B, 122B-A10B, 35B-A3B, 27B, 9B, 4B, 2B and 0.8B (DeepLearning.AI), led by a text, image and video model (AlternativeTo) and tagged Apache-2.0. The Qwen3.8 open weights reached Hugging Face in early August (repositories created 5 and 8 August) as two very different releases: Qwen3.8-27B under Apache 2.0 (6.8M downloads) and Qwen3.8-2.4T-A95B under a custom "Qwen3.8-Max License" (32.5k downloads). A third model, Qwen3.8-Flash-Next (180B parameters, 1.87M downloads), carries an "other" licence tag (Hugging Face API, Qwen). Lin Junyang, technical lead of the Qwen model team, resigned on 3 Mar 2026 (TechNode, citing people familiar; secondary).

Two widely repeated claims about Qwen are wrong, and the corrections matter commercially. A June blog said Alibaba had moved its flagship behind an API and charged a revenue share above US$50M (Jason Pollak blog). The licence text, read in full (and re-read on 11 October 2026), says otherwise. A licensee needs a separate licence before commercial use only if it or its affiliates run a "Model as a Service" or "AI Work Assistant" business and the aggregate revenue of the licensee and its affiliates exceeds US$50M (about ₹480 crore) in any 12 months. The licence defines Model as a Service as giving third parties inference or fine-tuning access with "meaningful control over the inputs, parameters, or training data", and an AI Work Assistant as a stand-alone coding or office-productivity product. The model name must be displayed above 100M monthly active users or US$20M monthly revenue. Internal use is exempt only if it does not make the software, its outputs or its capabilities available to any third party, and the text contains no revenue share (Qwen3.8-Max LICENSE). The API snapshot qwen3.8-max-0902 is priced the same as the open 2.4T checkpoint, but they should not be assumed to be the same model: no "Max" weights appear in the Qwen organisation listing, and OpenRouter's records score the open checkpoint at 39.9 and the Max snapshot at 45.4 (live check, 11 October 2026; OpenRouter models API).

The second correction concerns capability. An earlier reading of OpenRouter's chart put Qwen3.8 Max at 53.4, tied with Claude Fable 5.1 on the Artificial Analysis Intelligence Index. OpenRouter's model record gives 45.4 for qwen3.8-max-0902, and 53.4 is Fable 5.1's own value, so this report uses 45.4. Qwen3.8 Max does still appear among the leaders in OpenRouter's share-of-spend list for one task category (which category was not captured), next to Claude and GPT models (OpenRouter rankings).

Alibaba charges by region, and the region you pick is also a data-residency choice:

Source: Alibaba Cloud Model Studio pricing, updated 11 Oct 2026. Beijing input prices run about 15% to 78% below Singapore. Singapore offers a 90-day free quota, and the page notes a 50% batch discount for qwen3.8-max. Alibaba's table heads the qwen3.8-max output column "Chain of thought + answer", so thinking tokens bill at the output rate. The Moonshot, Z.ai, MiniMax and Xiaomi pages do not say how reasoning tokens bill. For code, the page lists only qwen3-coder-flash ($0.30 / $1.50 in Singapore); locally, Qwen3-Coder-30B-A3B at 8-bit decodes at 58.4 tokens per second on a DGX Spark (llama.cpp DGX Spark benchmarks).

Qwen's reach goes beyond its own API. The Qwen app had 234M users by May 2026, and AI Singapore's Sea-LION switched its base model from Llama to Qwen (both single-source, Wikipedia, Qwen). An April 2026 arXiv study credits Qwen's many sizes with giving it the broadest adoption of any open family (ATOM report; abstract only). The claim that 40% of new derivative models are Qwen-based rests on a single blog.

Moonshot Kimi K3 is a 2.78-trillion-parameter open model with a US$20M trigger

Moonshot's Kimi K2.5 (27 January 2026) is a multimodal mixture-of-experts model with about 1T total and 32B active parameters and an "Agent Swarm" mode of up to 100 parallel sub-agents (vendor and press description, NYU Shanghai RITS). Epoch ranked it 9th of 179 at release with an ECI of 148 (Epoch AI), while BenchLM placed it 39th of 88 (BenchLM). On the official SWE-bench bash-only board it scored 70.8% in February 2026 (SWE-bench leaderboard data).

Kimi K3 is confirmed as open weights. The Hugging Face repository holds 2.78T parameters in safetensors and has 1.13M downloads. OpenRouter lists it from 16 July 2026 as "a 2.8T parameter open-weight multimodal reasoning model" with a 1,048,576-token context (Hugging Face API, moonshotai; OpenRouter models API). The Hugging Face repository was created on 13 June, but Microsoft's Q2 2026 diffusion report says the full weights "were released in July", which fits the OpenRouter listing date (a Hacker News-listed item gives 27 July) (Microsoft AI Diffusion Report 2026 Q2). This report treats July 2026 as the release month. Moonshot's site now headlines "Kimi AI with K3 — Built for Agentic Coding & Knowledge Work" (kimi.com).

The licence was read in full (and re-read on 11 October 2026). If the licensee or an affiliate operates a hosted-model service ("Model as a Service") and the group's aggregate revenue exceeds US$20M (about ₹192 crore) in any 12 consecutive months, it needs a separate agreement before commercial use. The definition of Model as a Service excludes "end-user products with model capabilities solely embedded within specific features or harnesses" and the mere relaying of requests. "Kimi K3" must be displayed above 100M monthly users or US$20M monthly revenue. Internal use, defined as use that makes neither the software, its outputs nor its capabilities available to third parties, is exempt, as is use through Moonshot's own products or certified inference partners (Kimi-K3 LICENSE). Wikipedia says K3 "requires revenue sharing of up to 30%" from large inference providers (Wikipedia, Moonshot AI). That clause is not in the licence text, so the text wins.

The two price sources do not agree. Moonshot's undated platform page lists kimi-k3 at $0.30 cache-hit / $3.00 cache-miss input and $15.00 output, with cache writes at $3.00 (5-minute) or $6.00 (1-hour) (re-read on 11 and 12 October 2026), and the cheaper kimi-k2.7-code and kimi-k2.6 at $0.95 cache-miss / $4.00 output with a 262,144-token context (Kimi platform pricing). OpenRouter's listing varies between reads: $0.73–0.80 input and $10–13.50 output, with $0.80 / $0.55 / $13.50 on 11 Oct 2026 (OpenRouter models API). Moonshot's page ($3 / $0.30 / $15; context 1,048,576) is the list price; cite OpenRouter only as a dated listing. K3 scores 43.6 on Artificial Analysis's index (v4.3.2, 12 Oct 2026, primary; Artificial Analysis) and is #16 on LMArena's 8 Oct 2026 text snapshot (1488) (LMArena text). AWS Bedrock serves K3 through an India geography profile (in.moonshotai.kimi-k3) whose only destinations are Mumbai and Hyderabad (AWS Bedrock regional availability, primary). K2.6's 27.0 on OpenRouter's record of the index cannot be compared with older third-party figures, probably because the index version has changed.

Zhipu's GLM is the volume leader among cheap open weights

Z.ai (Zhipu) launched GLM-5, a 744B mixture-of-experts model under the MIT licence, on 11 February 2026 (Wikipedia gives 12 February) (NYU Shanghai RITS; Hugging Face API, zai-org). The same day, its GLM Coding Plan rose 30% for new subscribers (Techloy). On 7 April, the day it open-sourced GLM-5.1, it raised API prices by 10% (Wikipedia, Zhipu AI; other reports say 8% to 17%, Yahoo Finance). GLM-5.2 (MIT, 16 June) was reported as the top open-weight model on the Artificial Analysis index in June, but only one secondary source says so. GLM-5.3 and GLM-5.3-Flash followed on 25 August. The Flash model is MIT-licensed with 6.54M downloads, but the full GLM-5.3 is tagged "other", so the family is not uniformly MIT.

GLM's pull is price. Z.ai's own page lists GLM-5.3-Flash at $0.15 input, $0.03 cached and $0.50 output (OpenRouter matches). GLM-5 costs $1.00 / $3.20, GLM-5.1 through 5.3 cost $1.40 / $4.40, GLM-OCR costs $0.03 each way, and three older Flash models are free (Z.ai pricing). OpenRouter's listing for the full GLM-5.3 ($0.079 / $4.80 on 12 Oct 2026) differs from Z.ai's list price ($1.40 / $0.26 / $4.40); quote the vendor. The result is volume: GLM 5.3 Flash was second in OpenRouter's weekly token ranking (11.6T) and processed 55.1T tokens over 30 days, up 112%. Z.ai took 6.6% of OpenRouter text requests in the week beginning 28 September (OpenRouter rankings). On capability, GLM-5.3 scores 44.8 on Artificial Analysis's index (v4.3.2, 12 Oct 2026, primary; Artificial Analysis) and 41.82% on the Terminal-Bench 4.0 board (max effort; tbench.ai). GLM 5.3 Flash's 41.8 is OpenRouter's record (live check, 11 October 2026). Microsoft's Q2 2026 diffusion report, citing Artificial Analysis data, says GLM 5.3 Flash performs similarly to GPT-5.6 Sol at "high" reasoning for roughly a third of the cost (Microsoft AI Diffusion Report 2026 Q2). GLM 5 reached 72.8% on SWE-bench bash-only in February (SWE-bench leaderboard data).

Zhipu also publishes Open-AutoGLM, described as "An Open Phone Agent Model & Framework" (Apache-2.0, about 26.4k GitHub stars) (GitHub, Open-AutoGLM). Demand has outrun its capacity before: a late-February 2026 compute shortage forced a temporary sign-up restriction and a 23% share drop (Wikipedia, Zhipu AI).

MiniMax M3 is cheap but needs attribution from the first commercial use

MiniMax M3 appeared on OpenRouter on 31 May 2026. Its Hugging Face repository dates from 2 June and holds 427B parameters. OpenRouter lists a 1,048,576-token context and text, image and video input (Hugging Face API, MiniMaxAI; OpenRouter models API). It is not Apache. The MiniMax Community License grants rights for non-commercial purposes and requires that any commercial use, including commercial use of hosted APIs, display "Built with MiniMax M3". It also requires a one-time notice to MiniMax where the products or services using the model generate under US$20M a year, or prior written authorisation above that. "Commercial Use" explicitly covers offering paid products or services that rely on the model and deploying fine-tuned versions commercially. An appendix prohibits uses such as military purposes, harm to minors, discrimination, harmful misinformation and illegal content (MiniMax-M3 LICENSE, re-read on 11 October 2026). Together with Llama's attribution rule, this is the lowest bar for obligations in the report: the duties start with the first commercial use, and hosted-API use counts.

On price, the vendor page lists M3 at $0.60 / $2.40 up to 512K input tokens, marked "Permanent 50% off" to $0.30 / $1.20 (cache reads $0.06); above 512K the price doubles. M2.7, M2.5 and M2.1 cost the same (MiniMax pay-as-you-go pricing). MiniMax's coding record comes from the previous generation: M2.5 scored 75.8% on SWE-bench bash-only in February 2026, level with Gemini 3 Flash and close to the top Claude entry's 76.8% (SWE-bench leaderboard data). MiniMax's most-downloaded repository is not a chat model: MiniMax-H3 (33.1B parameters, 3.61M downloads, custom licence) generates video from images and text (MiniMax-H3 README). The company reported 2025 revenue of US$79.0M against an operating loss of US$1.87B (Wikipedia, MiniMax; single source).

Tencent brings Apache 2.0 to the 300B–780B class

Tencent publishes the largest Apache-licensed models in the Chinese set. Hy3 (2 July 2026) is Apache-2.0, with 298.8B parameters in safetensors and 496k downloads. OpenRouter describes it as a 295B mixture-of-experts model with 21B active, 192 experts and top-8 routing (Hugging Face API, tencent/Hy3). Hy4-preview (27 August) is also Apache-2.0, at 780B parameters (OpenRouter says 770B total, 49B active) and "designed for coding agents" (Hugging Face API, tencent/Hy4-preview). The earlier Hy3-preview and Hunyuan-A13B carry "other" licence tags, so check each repository. Tencent Cloud's own prices were not found. OpenRouter hosts list Hy3 at $0.083 / $0.33 and Hy4-preview at $0.751 / $2.251 (OpenRouter models API). Hy4-preview has only 25.7k downloads, yet it ranked fifth in OpenRouter's weekly tokens (8.41T, up 28%) and fourth over 30 days (40.6T) (OpenRouter rankings). Downloads and API traffic measure different things.

StepFun and Xiaomi rank high on OpenRouter but document little

StepFun (Shanghai, 2023; Tencent is an investor) released Step 3.5 Flash in February 2026: a 196B-total, 11B-active mixture-of-experts model under Apache 2.0 with a 256K context (Wikipedia, StepFun). Step-3.7-Flash (23 May 2026) keeps the Apache licence at 201.4B parameters and adds FP8 and NVFP4 variants. The NVFP4 build has 210k downloads (Hugging Face API, stepfun-ai). OpenRouter prices it at $0.20 / $1.15. Step 5 Preview was listed on OpenRouter on 8 October 2026 at $1.00 / $2.70 with a 1M context. It is described as a 600B-total, 27B-active sparse model and has no Hugging Face ID, so its open-weight status is unverified (OpenRouter models API). In its first days it was eighth by weekly tokens (6.37T) and second on the daily board (3.15T) (OpenRouter rankings).

Xiaomi is the least documented maker in this report. Its MiMo-V2.6-Flash was third in OpenRouter's weekly tokens (11.3T) and seventh over 30 days (25.1T). Xiaomi took 2.4% of text requests, up 16% in a week (OpenRouter rankings). Earlier sources gave no licence, size or price for MiMo; a check of the primary listings on 11 October 2026 found that Xiaomi's Hugging Face organisation holds MiMo-V2.6-Flash-RL (about 311B parameters) and MiMo-V2.6-Pro-RL (about 1.02T), both created on 21 September 2026 and tagged MIT (Hugging Face API, XiaomiMiMo). Xiaomi's own overseas price page lists MiMo-V2.6-Pro at $0.435 input (cache miss), $0.0036 cache hit and $0.87 output, and Flash at $0.14 / $0.28, with no input-length tiers since 27 May 2026 (Xiaomi MiMo pricing, vendor-reported; OpenRouter matches, with a 1.05M context; this Pro price is sometimes misattributed to DeepSeek V4-Pro). Artificial Analysis's own leaderboard gives Pro 46.3, the best open-weight score (v4.3.2, 12 Oct 2026, primary; Artificial Analysis); MiMo is not on the Terminal-Bench 4.0 board. The MIT tag sits only in the model-card metadata: the public Pro-RL and Flash-RL repos contain no LICENSE file, and the plain MiMo-V2.6-Pro and Flash repos are gated (HTTP 401) (MiMo-V2.6-Pro-RL). Treat MiMo as a serious candidate to test, but get written licence terms before any commercial use.

ByteDance and Baidu run mostly closed platforms at consumer scale

ByteDance's Doubao is the largest Chinese AI platform by usage. Volcano Engine said Doubao passed 120 trillion tokens a day in March 2026, double the figure three months earlier (TechNode; company-reported). Doubao had 330M users in May 2026 (Wikipedia, Doubao). One day of Doubao traffic exceeds the 101T tokens that OpenRouter's top model processed in 30 days, so router statistics miss most Chinese domestic usage. The Seed models are API-only; OpenRouter hosts list Seed 2.1 Turbo at $0.50 / $2.50 (262K context), Seed 2.0 Code at $0.50 / $3.00 and Seed 2.0 Mini at $0.10 / $0.40, and Volcano Engine's own price list could not be read (OpenRouter models API). ByteDance's open contribution is tooling: Coze Studio, an agent development platform, is Apache-2.0 with about 21.7k GitHub stars (GitHub, coze-studio).

Baidu's flagship is closed. ERNIE 5.0 (November 2025) is a proprietary, natively omni-modal model (VentureBeat), reported at 2.4T parameters. On 15 January 2026 it ranked 8th worldwide and first among Chinese models on LMArena, with 1,460 points (The Decoder). Wikipedia's infobox lists ERNIE 5.1 (29 April 2026) as Apache-licensed (Wikipedia, Ernie Bot), but no ERNIE 5.x weights appear on Hugging Face, so treat 5.x as closed. Baidu's open releases are the Apache-2.0 ERNIE 4.5 family (300B-A47B, 21B-A3B, and vision-language 424B-A47B and 28B-A3B) and document models: Unlimited-OCR (MIT, 1.22M downloads), Qianfan-OCR and ERNIE-Image (both Apache-2.0) (Hugging Face API, baidu). ERNIE API prices were not verified.

Western and other makers compete on provenance, platforms and small sizes

Outside China, the open-weight story in 2026 is mixed. One major lab moved its frontier behind closed doors. A European lab is about to publish a trillion-parameter model. Chip and enterprise-software makers (NVIDIA, IBM, Microsoft) have become serious open-model publishers at small and mid sizes. Company profiles are in Companies, and the regional view is in Americas and Europe.

Meta moved its frontier behind closed doors

Meta launched Muse Spark on 8 April 2026 as its first proprietary frontier model. Meta says it will keep releasing open versions, but those may lack some capabilities, and the largest models may stay closed. Late-2025 reporting said Meta had paused Llama 4 Behemoth, and Llama 4 itself drew "mixed and ultimately poor reviews" (all secondary, NYU Shanghai RITS). Muse Spark 1.1 followed on 9 July (Meta AI blog). Scale's SWE-bench Pro board (page dated 12 Oct 2026) is topped by Muse Spark 1.1 at 61.5% (asterisked; the asterisk is unexplained), ahead of GPT-5.4 (xHigh) at 59.1% (Scale SWE-bench Pro). OpenRouter lists Muse Spark 1.3 at $1.25 / $4.25 with a 1.05M context and an index score of 48.1, the highest outside the big five in this research (OpenRouter models API).

Llama is still downloaded heavily (Llama-3.2-1B has 6.9M downloads and Llama-3.1-8B 5.9M, per the Hugging Face API), and hosts sell it cheaply. OpenRouter lists Llama 4 Scout at $0.10 / $0.30 with a 1.31M context and Llama 4 Maverick at $0.1875 / $0.6525 (OpenRouter models API). The licence was read from Meta's GitHub copy. A licensee whose products exceeded 700M monthly active users in the month before release must request a licence, which Meta may refuse. Distributors must display "Built with Llama". Models built from Llama must start their names with "Llama". A notice text is required (Llama 4 Community License). The EU restriction on multimodal use is not in that copy and remains unverified. In practice, a Llama-based plan is a plan on an older generation whose next open release is uncertain.

Mistral bets on a trillion-parameter open release this month

Mistral's announcement of 6 October 2026 describes Mistral Large 4: 1T total and 52B active parameters, natively multimodal, in public preview API now, with "weights drop end of this month." It was trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centres. Mistral claims it is state of the art among open models in cyber, finance and law, and in the top five on the Artificial Analysis Cyber Index (vendor claims). The post does not name a licence (Mistral, Introducing Mistral Large 4). OpenRouter lists the preview at $0.68 / $2.09 with a 1,048,576-token context (OpenRouter models API). Arena's new-release board places it 42nd in WebDev (arena.ai leaderboard). Weights at the end of October are a stated plan, not a release. Check Hugging Face after 31 October before planning around them.

Mistral's licensing has an unresolved conflict. The pricing FAQ says open-weight models "are Apache 2.0 licensed for research/individual use; while commercial deployments require a Mistral license with separate terms for derivatives and production use" (Mistral pricing). Hugging Face metadata, however, tags Mistral-7B-Instruct-v0.3, Mistral-Small-3.1-24B and Voxtral-Mini-4B-Realtime as Apache-2.0. Until Mistral resolves this, read each model card and licence file before any commercial deployment. The FAQ gives Mistral Large (matching Large 3 on OpenRouter) at $0.50 / $1.50, with 50% off for batch and up to 90% off cached input. OpenRouter hosts list Medium 3.5 at $1.50 / $7.50, Small 4 at $0.15 / $0.60, Devstral at $0.40 / $2.00, and Ministral 3B, 8B and 14B at $0.10, $0.15 and $0.20 flat (OpenRouter models API).

Mistral also sells a platform: Workflows, Connectors with custom MCP tools, the Vibe agent, Forge, OCR 4 and open-weight Voxtral TTS all shipped in 2026. In September it raised a €3B Series D at a post-money valuation above €21B (Mistral news). Its pitch is sovereignty for European buyers, and it holds 2.0% of OpenRouter requests.

Microsoft pairs MIT-licensed Phi with its in-house MAI models

Microsoft's open line is small and permissive. Phi-4 is a 14B dense model with a 16K context under MIT, and Phi-4-mini-instruct has 3.8B parameters, a 128K context and an MIT licence (Phi-4 card; Phi-4-mini card). Downloads are modest (about 377k and 354k, Hugging Face API, microsoft), and OpenRouter lists Phi-4 at $0.07 / $0.14. Microsoft's in-house MAI family is offered through Foundry and Copilot, and no open weights for it were found. It began with image, voice and transcription models in Foundry in April 2026 and grew to seven models at Build 2026, led by MAI-Thinking-1 (the Build count is from secondary sources). Microsoft claims "top SWE-Bench Pro results" without publishing scores, and the reported 35B size and 256K context come from a single source (Microsoft AI news). MAI-Code-1.1-Flash is generally available in GitHub Copilot alongside Anthropic, OpenAI, Google and xAI models (GitHub Copilot supported models).

Amazon Nova 2 is a Bedrock-native agent stack with verified prices

Nova 2 Lite and Nova 2 Pro (preview) launched on Bedrock on 2 December 2025. Features include adjustable extended thinking, built-in code-interpreter and web-grounding tools, remote MCP tool support, a 1M-token context, and fine-tuning on Bedrock and SageMaker (AWS News summary). AWS's machine-readable price list settles conflicting third-party tracker figures. Nova 2 Lite costs $0.33 / $2.75 on demand in us-east-1, or $0.30 / $2.50 through the cross-region "global" profile, with batch at half price. Nova 2 Pro costs $1.375 / $11.00, or $1.25 / $10.00 global (AWS Bedrock price list). The AWS Nova page also lists Nova Act, "a new AWS service to build, deploy, and manage fleets of AI agents", and Nova Forge for building custom models on Nova (AWS Nova). Pro's general availability and Mumbai prices were not verified. From Mumbai, Nova's "Geo" profile is the APAC profile, which also routes to Tokyo, Seoul, Osaka, Singapore and Sydney, so it is not India-bound (AWS Bedrock regional availability). Check both before relying on Nova for Indian workloads.

Cohere sells sovereignty, not tokens

Cohere's pricing page lists no per-token prices. North (its agent platform) and Compass carry custom enterprise pricing. Model Vault instances are priced per hour or per month, for example Embed 5 Fast Small at $3.00 per hour or $2,000 per month, and Rerank 4 Pro Large at $10.00 per hour or $6,500 per month (Cohere pricing). Command A+ launched on 20 May 2026 (Cohere blog), and an OpenRouter host lists it at $0.30 / $1.50 with a 192,000-token context. Cohere Command A is offered on demand and on dedicated clusters in OCI's Hyderabad region, OCI's only India generative-AI region, together with Cohere Embed and Rerank. Llama 4 Maverick and Scout and gpt-oss-120b/20b are there on dedicated AI clusters only (OCI model regions, primary, 12 Oct 2026). Cohere and Aleph Alpha have signed a definitive business-combination agreement, subject to regulatory approvals and expected to close later in 2026 (no close was found by 12 Oct 2026), alongside a sovereign-AI partnership with Schwarz Group on STACKIT (Cohere and Aleph Alpha). An aggregator's roughly US$20B combined valuation is unverified; the last verified round was US$500M at US$6.8B in 2025 (Yahoo Finance).

IBM Granite and NVIDIA Nemotron are the enterprise-friendly open families

IBM positions Granite as the safe choice. Its page reads: "Apache 2.0 licensed. Cryptographically signed. ISO certified." The current Granite 4.2 language models come in 3B, 8B and 30B sizes (IBM Granite). IBM's footnote ties the certification to the "Granite AI Management System (AIMS)" under certificate 1102257-1 and does not name ISO/IEC 42001. Secondary claims that Granite was the "first" ISO 42001 family should therefore not be repeated as fact. Granite 4.0 (3 October 2025) introduced a hybrid Mamba-transformer design that, IBM says, runs on lower-cost GPUs (TechRepublic). OpenRouter hosts list Granite 4.2 8B at $0.06 / $0.25 with a 131,072-token context.

NVIDIA has become a major open-model publisher. The Nemotron 3 family uses a hybrid latent mixture-of-experts design, with Accenture, Deloitte, EY, ServiceNow, Palantir and Zoom named as early adopters (Enterprise AI World; secondary). Secondary sources sized Nemotron 3 Ultra (Computex, June 2026) at 500B or 550B; OpenRouter's model ID settles it at 550B total, 55B active, priced at $0.60 / $2.40. Super (120B / 12B) costs $0.08 / $0.45, and Nano (30B / 3B) $0.06 / $0.24 (OpenRouter models API). Nemotron 3.5 Lightning (30B mixture-of-experts, August 2026) targets always-on agents for code review, tool use and security-alert monitoring. It costs $0.07 / $0.20; the purpose is from secondary coverage (Pasquale Pillitteri). A secondary reconstruction placed Ultra near 48 on the Artificial Analysis index (Tech Insider), but OpenRouter's record of the index gives it 22.9 (live check, 11 October 2026; OpenRouter models API), so do not use the reconstruction. Ultra's free tier was seventh in OpenRouter's weekly tokens (6.63T), the only confirmed Western open model in the top ten (OpenRouter rankings). The NVIDIA Open Model License is described as permissive, but not every model card was read, so check them. A trillion-parameter Nemotron 4 is reported to be in development, from conflicting aggregator sources.

Apple, AI21, Liquid, Perplexity, Reka and Falcon fill niches

Apple opened its Foundation Models framework to third-party developers in iOS 27 (WWDC, 8 June 2026). One Swift API covers on-device and server-hosted inference (Techjack Solutions; secondary). Other reported details each rest on a single source: a protocol that lets one app call Apple's model, Claude or Gemini through the same code; free Private Cloud Compute for developers with fewer than two million first-time downloads (Origami); and a 4,096-token on-device context on A17 Pro or M1 hardware and newer (Kunal Ganglani). The models are reported to use Gemini-family technology (Callstack). They suit iOS-side summarising and extraction, not server pipelines.

Liquid AI's LFM2.5 line goes down to 230M parameters. The 2.6B GGUF build has 972,287 downloads (Hugging Face API, LiquidAI). Its licence is the one with the lowest cap in this report: commercial use "by a Legal Entity that exceeds the Threshold is not licensed", where the threshold is US$10M (about ₹96 crore) annual revenue (LFM Open License). AI21's small Jamba models (including Jamba-Reasoning-3B and Jamba2-3B) are tagged Apache-2.0 but have no OpenRouter listing (Hugging Face API, ai21labs). Perplexity sells search-grounded answers. OpenRouter lists Sonar at $1 / $1, Sonar Pro at $3 / $15 and Sonar Reasoning Pro at $2 / $8. Perplexity's own page now bills an "Orchestrator API" per token at each model's rate (Perplexity pricing). Reka Edge costs $0.10 / $0.10 with a 16,384-token context. TII's Falcon models with the most downloads date from 2023 to 2025, the newest of them Falcon-H1 (May 2025, licence tagged "other") (Hugging Face API, tiiuae). Sorted by date, TII's 2026 releases are small or specialised: the Falcon-H1-Tiny family (January 2026, 90M to 0.6B parameters by name, custom licence) and Falcon-OCR and Falcon-Perception (February 2026, Apache-2.0). Its newest 7B language model is Falcon-H1R-7B (October 2025) (Hugging Face API, tiiuae, newest first).

gpt-oss remains the open-weight reference point

No general-purpose open-weight release from OpenAI newer than the August 2025 pair was found. gpt-oss-120b has 117B parameters with 5.1B active and is described as fitting "into a single 80GB GPU". gpt-oss-20b has 21B with 3.6B active, for "lower latency, and local or specialized use cases". Both are Apache 2.0 (gpt-oss card). They have 3.7M and 5.9M downloads (Hugging Face API, openai). That is not the most among Western open models: Google's Gemma 4 26B-A4B (11.7M) and Meta's Llama-3.2-1B (6.9M) have more (Hugging Face API, google; Hugging Face API, meta-llama).

gpt-oss matters most as a price benchmark. Seventeen API providers serve gpt-oss-120b. The cheapest blended prices are CoreWeave at $0.04, DeepInfra at $0.05 and Crusoe at $0.07 per 1M tokens, and prices vary up to 8.9 times across providers. Output speeds reach 1,721 tokens per second on Cerebras and 473 on Groq (Artificial Analysis, gpt-oss-120b providers). The market already sells open models at close to cost. More on OpenAI's line is in OpenAI.

Small models for edge and on-device work

Small models are where licence clarity and hardware limits matter most. The sources reviewed confirm size and licence for a short list: Phi-4-mini (3.8B, MIT, 128K context), Granite 4.2 3B and 8B (Apache 2.0), Qwen3.5-4B and 9B (Apache 2.0 tags; 7.88M and 8.29M downloads), Liquid LFM2.5 (1.17B for the 1.2B-Instruct, revenue-capped) and AI21's Apache-tagged Jamba 3B models. Others have host prices and downloads only. Qwen3-0.6B has 30.6M downloads, the most of any repository in this research, followed by Qwen3-4B at 12.2M. OpenRouter hosts list Gemma 3 4B at $0.05 / $0.10, Llama 3.2 1B and 3B at $0.027 / $0.201 and $0.05 / $0.33, Ministral 3B and 8B at $0.10 and $0.15 flat, and Reka Edge at $0.10 / $0.10 with a 16,384-token context (OpenRouter models API; Hugging Face API, Qwen downloads). One tier up, gpt-oss-20b (3.6B active) and Nemotron 3 Nano (3B active) are server-class models with small active sizes. The sources reviewed contain no capability benchmarks and no measured on-device speeds for any of these, so treat them as a shortlist for testing, not a verdict.

The hardware floor for edge work is set by the platforms. Microsoft's Copilot+ PCs require an NPU of "more than 40 trillion operations per second" (Microsoft Learn). NVIDIA's Jetson AGX Thor module offers 128 GB of memory at 273 GB/s within 40–130 W, and the T4000 offers 64 GB (NVIDIA Jetson Thor). Most speed measurements are for laptop and desktop chips. An Apple M4 Pro with 273 GB/s decodes a 7B model at 4-bit at about 51 tokens per second, and an RTX 5090 laptop GPU at about 156 (llama.cpp CUDA scoreboard; llama.cpp Apple Silicon). The one shared edge measurement found is MLPerf Inference v6.1's "agentic edge" test on qwen3.6-27b (vendor submissions; single-stream latency per query, lower is better): Jetson AGX Thor 1,466 ms, RTX PRO 4500 2,616–2,758 ms, DGX Spark 3,808 ms, AMD Strix Halo 7,059 ms (MLPerf v6.1 results). The hardware research in this set (Hardware) concludes that edge inference suits offline or privacy-bound jobs (OCR, classification, speech, short extraction), while heavy reasoning belongs on a workstation or in the cloud. For a business, Phi-4-mini (MIT) and Granite 4.2 3B (Apache) are the cleanest licences to test first. Liquid's models are a good fit only where neither your business nor any company building on them for you crosses US$10M in revenue.

Licences decide how you may use and ship a model

Model quality changes every quarter. Licence terms decide which business models stay open. The table below summarises the licence texts read for this research. It is a planning aid, not legal advice, and each repository's licence file should go to counsel before a commercial launch.

Sources: the licence files and tags linked in the master table, plus Mistral pricing FAQ.

Four scenarios cover most of what a business adopting open models does. Internal use, meaning your own staff tools and prototypes, is exempt or unrestricted under most licences read. The Kimi and Qwen texts define it narrowly: nothing, including the model's outputs, may be made available to third parties, so outputs delivered to your customers are not internal use. The exceptions are MiniMax M3, whose base grant is non-commercial and whose duties attach to any commercial use, with no internal-use exemption; Liquid above US$10M revenue; and Mistral's FAQ position on commercial production use.

Deployments built for you, where your team or an implementation partner deploys open weights on your servers or private cloud, are straightforward under Apache and MIT. Under the custom licences, the thresholds count the revenue of "the licensee and its affiliates". Whether your business or the partner is the licensee in a given contract is a legal question for counsel, and this report does not answer it.

Offering a hosted endpoint is where the custom licences bite. Kimi K3 requires a separate agreement once a company that runs a hosted-model business passes US$20M of group revenue, and Qwen3.8-Max at US$50M. Kimi's text excludes features embedded in an end-user product from its hosted-model definition; whether a given deployment falls inside that exclusion is a question for counsel. MiniMax M3 requires a label and a notice from the first commercial use. Apache and MIT models carry no such terms. In this report's judgement, an Indian SME, even one serving US customers, is unlikely to reach the Kimi or Qwen thresholds, but the MiniMax display duty applies at any size.

Attribution in your product is a branding question as much as a legal one. Llama requires "Built with Llama" and a "Llama…" name for derived models. MiniMax requires "Built with MiniMax M3". Kimi and Qwen require display only at 100M users or US$20M monthly revenue.

The practical rule is short. For anything customer-facing, hosted or embedded, default to Apache 2.0 or MIT weights and keep a licence register per deployment. Use the custom-licence models where their capability is worth the paperwork.

Downloads favour small Qwen models; traffic favours cheap Flash models

There are two public adoption signals, and they measure different things. Hugging Face downloads show who builds on a model. The distribution channel itself is changing hands. On 2 Sep 2026 Nvidia signed a definitive agreement to buy Hugging Face for about $11.9B to stockholders, plus up to about $1.0B in retention awards, with closing expected in H1 2027 subject to approvals; Nvidia commits to keep the platform open (Nvidia Form 8-K, primary). The second signal, OpenRouter, shows which models developers route traffic to through one aggregator. OpenRouter itself warns that its token leaderboards and request-share chart count tokens and requests, and that "none of them measures users or spend"; only its per-task "Top models by task" lists are ranked by share of spend (OpenRouter rankings). The window behind the Hugging Face downloads field was not verified, so compare those counts only roughly.

Sources: Hugging Face API for Qwen, zai-org, MiniMaxAI, moonshotai, tencent, google, meta-llama, openai, baidu, XiaomiMiMo. The Qwen, Z.ai, MiniMax, Moonshot, Tencent and Xiaomi counts were re-read on 11 October 2026 and match. Lists are the top repositories per organisation, so a missing model is not proof of low use.

Source: OpenRouter rankings. By requests in the week beginning 28 September 2026, DeepSeek led with 21.8%, followed by Google (21.5%), OpenAI (18.2%), stealth models (10.5%), Z.ai (6.6%), Qwen (6.2%), Anthropic (2.5%), Xiaomi (2.4%) and Mistral (2.0%). The four Chinese authors in that list add up to 37.0% of requests. Tencent, StepFun, MiniMax and Moonshot fall below the top nine, so the Chinese total is a minimum. Spend may tell a different story. The one OpenRouter per-task share-of-spend list captured (its task was not recorded) was made up mostly of Claude and GPT models, headed by Claude, with Qwen3.8 Max the only Chinese entry (OpenRouter rankings). Read together, the two measures suggest that developers send high-volume, low-value traffic to cheap Chinese Flash tiers and pay for frontier closed models on the hard work. That reading is an inference from price and volume, not a claim made by the source. Microsoft's own analysis of public OpenRouter data puts open-weight models at 75% of OpenRouter token usage, while warning that these "are not global market shares" (Microsoft AI Diffusion Report 2026 Q2).

Capability evidence for these families is thin and comes in different formats. The cross-vendor figures available are below. Do not average or mix them.

Sources: Artificial Analysis leaderboard, OpenRouter models API, tbench.ai, Epoch AI benchmark data, SWE-bench leaderboard data, Epoch AI benchmarks, Epoch AI, Kimi K2.5, The Decoder, arena.ai leaderboard, LMArena text.

On this evidence, the best open-weight score (46.3) sits about 11 index points below the top closed model (57.6) and about 6 below GPT-6 Astra (52.7). In this report's judgement the gap is real but matters less for the high-volume, well-specified tasks (triage, extraction, routing, summarisation) that make up much of SME automation. See Use cases for which tasks those are.

Chinese-model risk is manageable with self-hosting, not with endpoints

Buyers outside China face five distinct risks, and the sources reviewed document them unevenly. The table states what the evidence actually shows and what to do about each risk.

India adds its own deadline. Under the DPDP Gazette notification of 13 November 2025, consent-manager provisions start on 13 November 2026. The core duties take effect on 13 May 2027: notice, consent, security safeguards, breach intimation, data-principal rights and the cross-border provisions (Gazette G.S.R. 843(E)). Rule 15 of the DPDP Rules lets personal data be transferred abroad "subject to" requirements the Central Government specifies by general or special order about making it available to a foreign State or its agencies. The power to restrict transfers to notified countries is s.16(1) of the Act. Both apply from 13 May 2027, and no s.16 notification or Rule 15 order was found. Separately, Rule 13(4) lets the government require Significant Data Fiduciaries to keep specified categories of personal data and traffic data in India, on a committee's recommendation (DPDP Rules, G.S.R. 846(E), primary). A business that processes Indian personal data through a PRC-hosted endpoint should still expect to document and justify that data flow once those provisions apply. That is an inference, not a stated rule. Self-hosting keeps the question simple. Sector rules: CERT-In's binding s.70B(6) Directions (28 Apr 2022) require listed cyber incidents, including attacks on AI and ML systems, to be reported within 6 hours, with logs kept for 180 days in India (CERT-In Directions, primary). RBI's FREE-AI framework (13 Aug 2025) is a committee report, not a direction (RBI press release). See Governance and India.

Which open model for which job and hardware tier

Two questions decide whether a model is usable for a given job: does it fit the hardware, and does it suit the task? The hardware tiers below combine sourced prices and measured speeds with this document's own memory arithmetic. The gpt-oss-120b file at 4-bit (MXFP4) takes 59.02 GiB for 116.8B parameters, about 0.54 bytes per parameter. Applying that ratio to other models gives approximate weight sizes. These estimates exclude the KV cache and runtime overhead, and quality at 4-bit was not tested for most models.

Sources: llama.cpp DGX Spark benchmarks, llama.cpp CUDA scoreboard, llama.cpp Apple Silicon, Lucebox, Strix Halo, NVIDIA DGX Spark, DGX Spark price report, 13 Aug 2026, NVIDIA DGX Station, NVIDIA RTX PRO 6000, NVIDIA RTX 5090, Framework mainboard, Apple Mac Studio specs, Apple store, Mac Studio, Apple India store, Mac Studio, Azure retail prices, Lambda pricing, ECorpIT, E2E Networks, E2E H100, IndiaAI price calculator, AWS EC2 price list, ap-south-1, Artificial Analysis. The hardware research finds that single-stream speed tracks memory bandwidth: DGX Spark (273 GB/s) is fast enough for a small team, not for many concurrent users. More detail is in Hardware.

Self-hosting rarely beats the market on cost alone. One H100 rented at Lambda's $3.99 an hour (its largest-instance rate; smaller instances cost $4.09–4.29) costs about $2,913 (≈₹2.8 lakh) a month. To beat gpt-oss-120b's $0.17 per 1M output tokens, it would need about 17 billion output tokens a month, roughly 6,500 tokens per second sustained at full utilisation, or about 7,000 at $4.29 (Lambda pricing; calculation in Costs). The costs analysis concludes that self-hosting pays off when there is sustained load, a data-residency rule or a fine-tuned model that no provider serves. Price per token alone is not the reason. See Costs.

The job guide below pairs common SME and enterprise automations with open-weight first choices and hosted alternatives.

Agent frameworks are covered in Frameworks and the wider stack in Tool stack.

The per-task arithmetic shows how small the model bill is once a job is routed to the right tier. The invoice-extraction profile from the costs analysis assumes 3,500 input and 800 output tokens per three-page text PDF. That profile is an assumption, not a measurement. Under it, 1,000 invoices cost:

Rupee values are computed from unrounded dollar costs at ₹96. For comparison, the costs analysis puts the same job on Claude Haiku 5.5 at $0.75 per 1,000 (about ₹72) (Costs, Claude). OCR for scanned documents, retries, human review and integration are excluded, and the costs analysis suggests multiplying model cost by 1.2 to 2 for a production budget, which is itself an assumption. The spread between the cheapest and dearest rows is more than 150 times. Choosing the tier has a far bigger effect on the bill than choosing between vendors in the same tier.

What this means for your business

Match models to jobs, not to one vendor

Plan on two layers. Permissively licensed open weights that you can self-host or buy from many hosts (gpt-oss-20b and gpt-oss-120b, Qwen3.8-27B, Granite 4.2, GLM-5.3-Flash) cover triage, extraction, retrieval and summarisation at a small fraction of frontier prices. One or two frontier APIs handle multi-step reasoning, coding agents and computer use, where the best open model still trails the top closed model by about 11 index points (Top models). Route each task to the cheapest tier that passes your own tests: the 150-fold spread in the invoice table matters more than the choice between vendors in a tier.

What to budget

Model tokens are a small line: 1,000 invoices cost about ₹13 to ₹2,160 depending on the model, before OCR, retries, human review and integration (the costs analysis suggests multiplying model cost by 1.2 to 2). Do not self-host for price alone. Hosted gpt-oss-120b sells at about $0.04 blended, while a rented H100 costs about ₹2.8–3.0 lakh a month at Lambda's rates. Self-host when data rules require it, when sustained volume justifies it, or when a fine-tuned model has no hosted equivalent. A 128 GB desktop box (a DGX Spark at about ₹4.5 lakh at its US list price, or a Ryzen AI Max+ 395 board from about ₹3 lakh before case, power supply and storage) suits a small-team pilot, not many concurrent users.

Data handling and India processing

Separate where a model comes from from where your data goes. For personal or regulated data, the defensible default is to self-host weights in your own region or in India. Chinese Apache-licensed models qualify if your business accepts them, and gpt-oss, Granite, Nemotron and Phi are the Western-origin alternatives. Keep personal data off PRC-hosted endpoints; for non-personal work on Chinese hosted APIs, prefer Alibaba's Singapore region or a non-PRC host, and record the choice. Where processing must stay in India, AWS Bedrock's India profile for Kimi K3 and its In-Region Mumbai list of older open models (Qwen3, GLM 4.7/5, MiniMax M2.x, Kimi K2.5, Mistral Large 3, gpt-oss) are documented options. Mumbai prices were not captured, and AWS warns that prompts may be stored in opt-in Regions for abuse detection (Hyderabad is one, so still in India) (AWS Bedrock regional availability). The DPDP core duties start on 13 May 2027, so document these data flows now.

Licences and volatility

Keep a register of model, licence file version, deployment owner and whether the deployment is offered to others as a hosted service. Flag MiniMax M3 (a label and a notice from the first commercial use), Llama (naming and "Built with Llama"), Liquid LFM (a US$10M revenue cap that a growing business can cross), Mistral (the FAQ conflict) and Xiaomi MiMo (MIT only in model-card metadata; no LICENSE file in the public repos). The Kimi K3 and Qwen3.8-Max thresholds are far above SME scale. None of this replaces counsel. Chinese vendors changed prices, capacity and terms several times in 2026, and new models arrive monthly, so pin dated snapshots, keep a second model behind the same interface for each workflow, and re-run your evaluation suite whenever a model is swapped (Playbook).

What to test

Test task quality on your own data, because OpenRouter shows where developers send cheap traffic, not what enterprises buy. Also test refusals and bias on your own prompts, Indian-language quality (no independent 2026 benchmark covers these models apart from one preprint), quality at 4-bit if you self-host, and actual bills on Chinese APIs that do not say how reasoning tokens bill.

Questions to ask any vendor or implementation partner

• Which model and licence file version will be used, and who is the licensee under the contract?
• Where does each request run, and how was the route verified (on Bedrock, with GetInferenceProfile)?
• What is the model's origin, and is a Western-origin alternative ready if needed?
• Which dated snapshot is pinned, and what is the fallback?
• What results can they show on your own data, rather than on public leaderboards?

Open questions and unverified items

Sources

1. Hugging Face API, Qwen organisation (by last modified)
2. Qwen3.8-Max License (Qwen3.8-2.4T-A95B)
3. Kimi K3 License
4. MiniMax M3 Community License
5. Hugging Face API, zai-org
6. Hugging Face API, tencent/Hy3
7. Hugging Face API, tencent/Hy4-preview
8. Hugging Face API, stepfun-ai/Step-3.7-Flash
9. Llama 4 Community License (Meta GitHub copy)
10. Liquid AI LFM Open License v1.0
11. gpt-oss-20b model card
12. Microsoft Phi-4 model card
13. Microsoft Phi-4-mini-instruct model card
14. Alibaba Cloud Model Studio pricing
15. Kimi platform pricing
16. Z.ai pricing
17. MiniMax pay-as-you-go pricing
18. AWS Bedrock price list, us-east-1
19. OpenRouter models API
20. OpenRouter rankings
21. Mistral, Introducing Mistral Large 4
22. IBM Granite
23. Madrimasd, Zhipu and MiniMax Hong Kong debut
24. AIWeekly, MiniMax and Zhipu enter Hang Seng Tech Index
25. Wikipedia, Nvidia
26. DeepLearning.AI, Qwen3.5 open-weight sizes
27. AlternativeTo, Qwen3.5 launch
28. Jason Pollak blog, China AI models 2026 (secondary)
29. llama.cpp DGX Spark benchmark file
30. Wikipedia, Qwen
31. ATOM project report, arXiv 2604.07190
32. NYU Shanghai RITS, Kimi K2.5
33. Epoch AI, Kimi K2.5
34. BenchLM, Kimi K2.5
35. SWE-bench official leaderboard data
36. Hugging Face API, moonshotai
37. kimi.com
38. Wikipedia, Moonshot AI
39. NYU Shanghai RITS, GLM-5
40. Techloy, GLM-5 launch and price increase
41. Wikipedia, Zhipu AI
42. Yahoo Finance, Zhipu raises AI model prices
43. GitHub, Open-AutoGLM
44. Hugging Face API, MiniMaxAI
45. MiniMax-H3 README
46. Wikipedia, MiniMax (company)
47. Hugging Face API, stepfun-ai
48. Wikipedia, StepFun
49. TechNode, Doubao passes 120 trillion daily tokens
50. Wikipedia, Doubao
51. GitHub, Coze Studio
52. VentureBeat, Baidu ERNIE 5
53. The Decoder, ERNIE 5.0 on LMArena
54. Wikipedia, Ernie Bot
55. Hugging Face API, baidu
56. NYU Shanghai RITS, Meta Muse Spark analysis
57. Meta AI blog
58. arena.ai leaderboard
59. Mistral pricing and FAQ
60. Mistral news
61. Hugging Face API, microsoft
62. Microsoft AI news
63. GitHub Copilot supported AI models
64. AWS News summary, Nova 2 on Bedrock
65. AWS, Amazon Nova
66. Cohere pricing
67. Cohere blog
68. Cohere, Cohere and Aleph Alpha sign agreement
69. Yahoo Finance, Cohere valued at $6.8B
70. TechRepublic, IBM Granite 4.0
71. Enterprise AI World, NVIDIA Nemotron 3 family
72. Pasquale Pillitteri, Nemotron 3 Ultra
73. Tech Insider, Nemotron 3 Ultra
74. Techjack Solutions, Apple Foundation Models API
75. Origami, Apple Foundation Models
76. Kunal Ganglani, Apple Foundation Models for developers
77. Callstack, On-device AI after WWDC 2026
78. Hugging Face API, LiquidAI
79. Hugging Face API, ai21labs
80. Perplexity API pricing
81. Hugging Face API, tiiuae
82. Hugging Face API, openai
83. Hugging Face API, google
84. Hugging Face API, meta-llama
85. Artificial Analysis, gpt-oss-120b providers
86. Hugging Face API, Qwen organisation (by downloads)
87. Microsoft Learn, Copilot+ PCs developer guide
88. NVIDIA Jetson Thor
89. llama.cpp CUDA scoreboard
90. llama.cpp Apple Silicon benchmarks
91. Hugging Face API, tencent
92. Epoch AI benchmarks hub
93. Wikipedia, Cambricon Technologies
94. Wikipedia, iFlytek Spark
95. Wikipedia, Manus (AI agent)
96. Digitimes, Manus probe, 18 Mar 2026
97. Kimi on X, subscription pause post (not read)
98. Gazette of India, DPDP commencement G.S.R. 843(E)
99. Lucebox, DeepSeek V4 Flash on Strix Halo
100. NVIDIA DGX Spark
101. ainauten, DGX Spark price rise, 13 Aug 2026
102. NVIDIA DGX Station
103. NVIDIA RTX PRO 6000 Blackwell
104. NVIDIA RTX 5090
105. Framework Desktop mainboard, Ryzen AI Max
106. Apple Mac Studio specifications
107. Azure Retail Prices API, ND96isr H100 v5
108. Lambda pricing
109. ECorpIT, India GPU cloud rental pricing 2026
110. E2E Networks, H100 pricing in India
111. Bloomberg, Z.ai's Ox Alpha stealth model (listed on Hacker News)
112. Microsoft AI Diffusion Report 2026 Q2 (Sept 2026)
113. Hugging Face API, XiaomiMiMo (read 11 Oct 2026)
114. Gazette of India, DPDP Rules 2025, G.S.R. 846(E) (13 Nov 2025)
115. Hugging Face API, tiiuae, newest first
116. Artificial Analysis, model leaderboard, Intelligence Index v4.3.2 (live 12 Oct 2026)
117. Terminal-Bench 4.0 board (tbench.ai)
118. LMArena text leaderboard (8 Oct 2026 snapshot)
119. Scale, SWE-bench Pro leaderboard (12 Oct 2026)
120. Epoch AI, benchmark data download
121. Xiaomi MiMo pay-as-you-go pricing
122. Hugging Face, XiaomiMiMo/MiMo-V2.6-Pro-RL
123. TechNode, Qwen technical lead Lin Junyang leaves Alibaba (4 Mar 2026)
124. AWS Bedrock, model support by Region (regional availability)
125. Oracle OCI Generative AI, models by region
126. E2E Networks, inference endpoints (vendor)
127. Yotta Shakti Cloud, AI endpoints (vendor)
128. CERT-In Directions under s.70B(6) (28 Apr 2022)
129. CERT-In, SBOM/QBOM/CBOM/AIBOM/HBOM technical guidelines v2.0
130. RBI press release, FREE-AI committee report (13 Aug 2025)
131. MLPerf Inference v6.1 results, summary.csv
132. vLLM blog, Qwen3.5-397B at 25K tok/s per GPU (6 Aug 2026)
133. vLLM blog, MiniMax M3 on MI355X (10 Sep 2026)
134. Indi-RomCoM preprint, arXiv 2606.30790
135. Hugging Face, Gnani gnani-evon-v3.3-30B-A3B model card (vendor-run MILU table)
136. Nvidia Form 8-K, Hugging Face agreement (filed 3 Sep 2026)
137. Apple store, Mac Studio (US)
138. Apple India store, Mac Studio
139. E2E Networks, NVIDIA H100 pricing
140. IndiaAI compute portal, price calculator
141. AWS Price List, EC2 ap-south-1 (Mumbai)