The largest open-weight models now reach hundreds of billions to trillions of parameters, but their licences, memory needs, serving engines and reasoning controls differ so much that "open weights" alone says little about whether a team can legally and practically run one. This page compares the largest current open-weight language model from each of twelve providers, using identical fields: licence and its stated conditions, total and active parameters, architecture, context length, input modalities, published precision formats, named inference engines, stated hardware, reasoning controls and whether the provider also sells the model as a hosted API. It is written for platform teams, ML engineers, legal reviewers and buyers deciding whether to self-host. For the basic open versus closed distinction, read Open-Weight vs Closed Models Explained first. Providers are listed alphabetically; the order is never a ranking, and the page draws no verdicts.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. Model cards and licences change between releases, so read the licence file in the repository before use.
Key takeaways
- Licences fall into three groups. Five models use a standard licence (Apache 2.0 for Gemma 4, gpt-oss-120b, Hy3 and K-EXAONE 2.0; MIT for DeepSeek-V4-Pro), one uses OpenMDW-1.1, and six use a publisher's own licence.
- Size spans roughly two orders of magnitude. Total parameters run from 31B (Gemma 4 31B) to 2.8T (Kimi K3.1 Pro), and active parameters per token from 5.1B (gpt-oss-120b) to 104B (Kimi K3.1 Pro).
- Hardware statements are rare. Only Google, Meta, NVIDIA, OpenAI and Tencent state memory or GPU needs on the pages checked; for the rest, the serving examples are the only guide.
- Low-precision weights are now published by providers. DeepSeek (FP4 plus FP8), Meta (NVFP4), Moonshot AI (MXFP4, trained with quantisation-aware training), NVIDIA (NVFP4), Tencent (FP8) and Z.ai (FP8) publish reduced-precision checkpoints or formats.
- Most also run as hosted APIs. Alibaba, DeepSeek, Mistral AI, Moonshot AI, OpenAI (including Amazon Bedrock) and Z.ai document hosted versions, so the same model can be tested by API before committing to hardware.
Which models are included?
The inclusion rule is objective: for each provider, the largest current open-weight text-generation model whose model card or official documentation could be fetched on 2026-11-15. Preview checkpoints are excluded (for example Tencent Hy4-preview and StepFun Step-5-Preview), as are providers whose largest open release is under 30B parameters. Each provider's smaller open models, such as Mistral Small 4, Qwen3.8-27B or Nemotron 3 Super, are covered on the provider pages. Restricted-access status is stated where a card requires an access request.
Open-weight frontier models at a glance
As of 2026-11-15. Parameters are the figures in the model card text; where the Hugging Face file header differs, the card text is used and the difference is noted below the table.
| Provider | Model | Total / active parameters | Architecture (as stated) | Context length | Input modalities | Licence |
|---|---|---|---|---|---|---|
| Alibaba (Qwen) | Qwen3.8-2.4T-A95B | 2.4T / 95B | MoE, 512 experts (10 routed, 1 shared); Gated DeltaNet and Gated Attention hybrid | 262,144 native, up to 1,010,000 | Text | qwen3.8-max (publisher licence) |
| DeepSeek | DeepSeek-V4-Pro | 1.6T / 49B | MoE; Compressed Sparse Attention and Heavily Compressed Attention | 1M | Text | MIT |
| Gemma 4 31B | 31B (dense) | Dense; the family also has a 26B A4B MoE | 256K (medium models) | Text, image (video and audio support vary by size) | Apache 2.0 | |
| LG AI Research | K-EXAONE 2.0 750B-A37B | 750B / 37B | MoE, 256 experts (8 active, 1 shared); sliding-window and global attention | 262,144 | Text | Apache 2.0 |
| Meta | Muse Spark Open 816B-A42B | 816B / 42B | MoE, 256 experts (8 active), with a 1.8B Perception Encoder | 1M | Text, image, video, audio | Muse Model License (publisher licence) |
| MiniMax | MiniMax-M3 | About 428B / about 23B | MoE with MiniMax Sparse Attention | 1M | Text, image, video | minimax-community (publisher licence) |
| Mistral AI | Mistral Large 4 | 1.05T / 52B, plus a 1.6B vision encoder | Granular MoE | 1M | Text, image | Mistral Ecosystem Open Weight License (publisher licence) |
| Moonshot AI | Kimi K3.1 Pro | 2.8T / 104B | MoE, 896 experts (16 selected, 2 shared); KDA and Gated DSA attention; 401M MoonViT-V2 vision encoder | 1,048,576 | Text, image | Kimi K3.1 Pro License (publisher licence) |
| NVIDIA | Nemotron 3 Ultra 550B-A55B | 550B / 55B | LatentMoE hybrid of Mamba-2, MoE and attention layers | Up to 1M (256K default serving configuration) | Text | OpenMDW-1.1 |
| OpenAI | gpt-oss-120b | 117B / 5.1B | Not stated by OpenAI on the model page beyond total and active parameters | 131,072 | Text | Apache 2.0 |
| Tencent | Hy3 | 295B / 21B (plus 3.8B MTP layer) | MoE, 192 experts (top 8) | 256K | Text | Apache 2.0 |
| Z.ai | GLM-5.3 | 753B / not stated by Z.ai on the card | Tagged glm_moe_dsa; shares its base model with GLM-5.2 | Not stated as a specification on the model card | Text | glm-5.3 (publisher licence) |
File headers on Hugging Face differ from card text for Meta (822B), NVIDIA (561B), Tencent (299B) and LG AI Research (749B). Google's Gemma 4 page groups sizes into small models with 128K context and medium models with 256K, and lists native audio input on E2B, E4B and 12B only. For how total and active parameters relate, see Mixture of Experts Explained.
What does each licence allow?
Licence names come from each model card. Conditions are listed only where the card or official page states them; where the card names a licence without summarising it, the licence file in the repository is the only source of the terms.
| Provider | Licence | Conditions stated on the page checked | Access |
|---|---|---|---|
| Alibaba (Qwen) | qwen3.8-max | Terms not summarised on the model card | Public repository |
| DeepSeek | MIT | The card states the repository and the model weights are licensed under the MIT License | Public repository |
| Apache 2.0 (linked as the Gemma 4 license) | Google states the models are provided with open weights and permit responsible commercial use | Downloads on Kaggle and Hugging Face | |
| LG AI Research | Apache 2.0 | Released so the community can inspect, deploy and build on the model | Public repository |
| Meta | Muse Model License (dated 15 October 2026) | Display "Built with Muse Spark Open"; a licensee offering AI models as a service by API with aggregate gross revenue above US$20 million in a 12-month period must notify Meta and sign a separate agreement first (internal use, research and features in the licensee's own products are excluded); usage policy in Exhibit A; not intended for users under 18 | Gated: an access request form asks for legal name, date of birth and organisation |
| MiniMax | minimax-community | Terms not summarised on the model card | Public repository |
| Mistral AI | Mistral Ecosystem Open Weight License | Terms not summarised on the model card | Hugging Face repository linked from the card |
| Moonshot AI | Kimi K3.1 Pro License | Terms not summarised on the model card | Public repository |
| NVIDIA | OpenMDW-1.1 | The card states the model is ready for commercial and non-commercial use | Public repository; some training datasets gated |
| OpenAI | Apache 2.0 | Described as a permissive Apache 2.0 licence | Weights on Hugging Face |
| Tencent | Apache 2.0 | No territory restriction stated on the card | Public repository |
| Z.ai | glm-5.3 | Terms not summarised on the model card | Public repository |
As of 2026-11-15. Licences can change between versions of the same family, so record the licence name, version and repository commit for every checkpoint you deploy. For approval workflows see the Open Source Licence Compliance Guide.
How much hardware do they need?
Hardware figures below are only those stated by the provider. Where no figure is stated, the provider's own serving example is described; examples are not minimums.
| Provider | Model | Published precision formats | Hardware stated or shown |
|---|---|---|---|
| Alibaba (Qwen) | Qwen3.8-2.4T-A95B | BF16 | No requirement stated; Docker example uses all GPUs and 32 GB shared memory |
| DeepSeek | DeepSeek-V4-Pro | FP4 (MoE experts) plus FP8 mixed; Base checkpoint FP8 mixed | No requirement stated; SGLang example uses all GPUs and 32 GB shared memory |
| Gemma 4 31B | BF16, 8-bit, 4-bit (Q4_0) memory figures | 69.9 GB (BF16), 34.9 GB (8-bit), 17.5 GB (Q4_0) to load weights; the 26B A4B needs 57.7 GB, 28.8 GB and 14.4 GB; figures exclude KV cache | |
| LG AI Research | K-EXAONE 2.0 | BF16 | Examples run on two nodes of 8 NVIDIA H200 GPUs with tensor parallelism of 16; a B200 flag is needed in SGLang to avoid a known generation issue |
| Meta | Muse Spark Open | BF16, NVFP4 | At least 700 GB VRAM for NVFP4; about 2 TB for BF16; optimised for two Mac Studio M5 Ultra machines with MLX |
| MiniMax | MiniMax-M3 | BF16; quantised versions linked | No requirement stated; SGLang example uses all GPUs and 32 GB shared memory |
| Mistral AI | Mistral Large 4 | Not stated on the model card | Not stated on the model card |
| Moonshot AI | Kimi K3.1 Pro | MXFP4 weights and MXFP8 activations (quantisation-aware training) | Not stated on the model card |
| NVIDIA | Nemotron 3 Ultra | BF16; separate NVFP4 repository | BF16: 8 GB200, B200, GB300 or B300, or 16 H100, or 8 H200 |
| OpenAI | gpt-oss-120b | Not stated on the model page | "Fits into a single H100 GPU" |
| Tencent | Hy3 | BF16; separate Hy3-FP8 repository | For 8-GPU serving, H20-3e or GPUs with larger memory recommended |
| Z.ai | GLM-5.3 | BF16, FP8 | No VRAM figure; Ascend NPU support through vLLM-Ascend, xLLM and SGLang |
As of 2026-11-15. Weight memory is only part of the total: the KV cache grows with context length and concurrent requests, which is why long-context serving often needs more GPUs than the weights alone suggest. See Running LLMs Locally Explained for the smaller end of the range.
Which inference engines are named?
| Provider | Engines named on the card | Quantised community builds mentioned |
|---|---|---|
| Alibaba (Qwen) | Transformers, vLLM, SGLang, TokenSpeed, Docker Model Runner | llama.cpp, Ollama, LM Studio, Jan |
| DeepSeek | vLLM, SGLang, Transformers, Docker Model Runner | llama.cpp, Ollama, LM Studio, Jan |
| Not listed on the docs page checked; downloads on Kaggle and Hugging Face | Q4_0 memory figures published | |
| LG AI Research | SGLang and vLLM (LG AI Research forks), Transformers, Docker Model Runner | llama.cpp, Ollama, LM Studio |
| Meta | SGLang, vLLM, Transformers, TensorRT-LLM, Unsloth, llama.cpp | llama.cpp, Ollama, LM Studio |
| MiniMax | SGLang, vLLM, Transformers, KTransformers, Unsloth, ATOM, Docker Model Runner | llama.cpp, Ollama, LM Studio |
| Mistral AI | Not stated on the model card | Not stated on the model card |
| Moonshot AI | vLLM, SGLang, TokenSpeed | Not stated on the model card |
| NVIDIA | vLLM, SGLang, TensorRT-LLM (Blackwell only at present), Transformers, Docker Model Runner | GGUF for llama.cpp, Ollama, LM Studio |
| OpenAI | Not stated on the model page; also served on OpenAI API endpoints | Not stated on the model page |
| Tencent | vLLM (MTP speculative decoding), SGLang (EAGLE speculative decoding), Transformers, Docker Model Runner | llama.cpp, Ollama, LM Studio |
| Z.ai | SGLang, vLLM, TokenSpeed, Transformers, KTransformers, Unsloth, vLLM-Ascend, xLLM, Docker Model Runner | llama.cpp, Ollama, LM Studio |
As of 2026-11-15. Inference servers expose network APIs and often run with broad GPU host access; see the vLLM, Triton and NVIDIA NIM Inference Server Security Guide.
Reasoning controls in the open checkpoints
Most of these models reason by default, and the control is usually a chat-template argument rather than an API field, so it must be set in the serving layer.
| Provider | Control | Values | Default |
|---|---|---|---|
| Alibaba (Qwen) | reasoning_effort | low, medium, xhigh; thinking cannot be disabled | xhigh |
| DeepSeek | Mode via prompt and template | Non-think, Think High, Think Max (at least 384K context recommended for Think Max) | Not stated on the model card |
| Configurable thinking modes | Not listed on the docs page checked | Not stated on the docs page checked | |
| LG AI Research | enable_thinking; preserve_thinking | True or False | Thinking on; preserve_thinking off (True recommended for agentic tasks) |
| Meta | reasoning_effort | low, medium, high, xhigh | xhigh |
| MiniMax | thinking | enabled, adaptive, disabled | Not stated on the model card |
| Mistral AI | reasoning_effort (API) | low, medium, high | high recommended |
| Moonshot AI | reasoning_effort | low, high, max | max |
| NVIDIA | enable_thinking; medium_effort; client-side reasoning_budget | On or off, medium effort, token budget | Thinking on |
| OpenAI | Reasoning effort | low, medium, high | Not stated on the model page |
| Tencent | reasoning_effort | no_think, low, high | no_think |
| Z.ai | reasoning_effort; clear_thinking template flag | low, high, max | max (clear_thinking false by default; set true for chat) |
As of 2026-11-15. For how these controls compare with closed-model APIs, see Reasoning LLMs Compared Across Providers.
Where can each model be used without self-hosting?
| Provider | Hosted option documented | Listed price (USD per 1M tokens, input / output) |
|---|---|---|
| Alibaba (Qwen) | Qwen3.8-Max on Qwen Cloud and Model Studio is the managed version of the checkpoint | $2 / $6 (international, Singapore) |
| DeepSeek | deepseek-v4-pro (DeepSeek-V4-Pro-0813) on the DeepSeek API; 1M context, 384K maximum output | $1.32 / $3.96 peak; $0.66 / $1.98 off-peak (cache miss) |
| Not stated on the Gemma docs page checked | Not stated on the Gemma docs page checked | |
| LG AI Research | No Hugging Face Inference Provider at the time checked; a FriendliAI API is linked from the card | Not stated on the model card |
| Meta | Not stated on the model card | Not stated on the model card |
| MiniMax | Not stated on the model card | Not stated on the model card |
| Mistral AI | mistral-large-4 on the Mistral API | $0.70 / $2.10 ($0.07 cached input) |
| Moonshot AI | kimi-k3.1-pro on the Kimi platform with OpenAI- and Anthropic-compatible interfaces; also HuggingChat and Inference Providers | Not stated by Moonshot AI on the pages checked (2026-11-15) |
| NVIDIA | Not stated on the model card section read | Not stated on the model card section read |
| OpenAI | OpenAI API endpoints and Amazon Bedrock (openai.gpt-oss-120b) | Not stated on the model page |
| Tencent | Not stated on the model card | Not stated on the model card |
| Z.ai | glm-5.3 on the Z.ai API | $1.40 / $4.40 ($0.26 cached input) |
As of 2026-11-15. The DeepSeek-V4-Pro model card describes the April 2026 release, while the API serves the DeepSeek-V4-Pro-0813 update; check which checkpoint matches the hosted version before comparing results. For the trade-offs between the two routes, see Self-Hosted LLMs vs Hosted LLM APIs.
Training data and transparency
- NVIDIA: the Nemotron 3 Ultra card states that major portions of the pre-training corpus, the post-training corpus and the reinforcement learning environments are released, with code, maths and multilingual data gated behind approval and some third-party datasets kept private; training recipes and a technical report are published.
- Other providers: the remaining cards checked describe architecture and serving and link technical reports or papers where they exist (DeepSeek, LG AI Research and MiniMax link papers); none of them describes a training data release.
- Base checkpoints: DeepSeek publishes a Base model alongside the instruct model (DeepSeek-V4-Pro-Base), which matters for teams that plan to fine-tune.
For what a model card should contain, see Model Cards and System Cards Explained.
Benchmarks
Each card publishes its own evaluation tables, with different test sets, effort levels and sampling settings (for example temperature 1.0 and top_p 1.0 on DeepSeek and Meta, top_p 0.95 on Kimi, MiniMax and LG AI Research, temperature 0.9 on Tencent). Because no single dated source covers every row, this page lists no scores. Compare scores only when the effort level, context and sampling settings match. See How to Read LLM Benchmarks.
Security and governance when self-hosting
- Supply chain: pull weights only from the publisher's own organisation on the hub, pin the repository commit and verify file hashes; see the Hugging Face Security Guide.
- Gated access: Meta's gated repository requires personal details on request; use an organisational account and record who accepted the licence.
- Serving layer: chat-template flags such as Z.ai's clear_thinking and NVIDIA's enable_thinking are set by whoever controls the server, so treat template files as configuration under change control.
- Safety tuning: once weights are local, provider-side filters do not apply; see the Open-Weight Model Security Guide and the Self-Hosted LLM Security Guide.
How to choose an open-weight model: factual criteria
| If you need | Check | Models documented with it |
|---|---|---|
| A standard licence your legal team already approves | Apache 2.0 or MIT | DeepSeek-V4-Pro (MIT); Gemma 4, gpt-oss-120b, Hy3, K-EXAONE 2.0 (Apache 2.0) |
| A single-GPU deployment | Provider hardware statement | gpt-oss-120b (single H100, per OpenAI); Gemma 4 31B at 8-bit or 4-bit by Google's memory table |
| Image, video or audio input | Input modalities | Muse Spark Open (text, image, video, audio); MiniMax-M3 (text, image, video); Gemma 4 (text and image, with audio on smaller sizes); Kimi K3.1 Pro and Mistral Large 4 (text, image) |
| Context of about 1M tokens | Stated context length | DeepSeek-V4-Pro, Kimi K3.1 Pro, MiniMax-M3, Mistral Large 4, Muse Spark Open, Nemotron 3 Ultra (with configuration), Qwen3.8-2.4T-A95B (extended) |
| Non-NVIDIA accelerators | Named hardware support | GLM-5.3 (Ascend NPU); Muse Spark Open (Apple MLX) |
| A hosted API for trial before self-hosting | Hosted option table | Qwen3.8-Max, deepseek-v4-pro, gpt-oss-120b, kimi-k3.1-pro, mistral-large-4, glm-5.3 |
Small teams usually start with models that fit one or two GPUs and a standard licence; at the trillion-parameter end, multi-node serving and a licence review are both part of the project plan. Larger organisations often pair a hosted API for evaluation with a self-hosted deployment once usage and licence terms are settled.
Practitioner checklist
- Read the licence file in the repository, not the licence tag, and record name, version and commit for each deployed checkpoint.
- Check revenue or service thresholds such as Meta's US$20 million API-service clause before offering a model to customers.
- Size hardware for weights plus KV cache at your target context and concurrency, not for weights alone.
- Prefer provider-published low-precision checkpoints (FP8, NVFP4, MXFP4) over community conversions where the provider offers them.
- Set the reasoning default explicitly in the chat template; several models default to their highest effort.
- Test the hosted API version first where one exists, and confirm it matches the checkpoint version you plan to run.
- Pin model repositories and verify hashes, and keep chat templates under change control.
- Re-check each card when a new version ships; parameters, licences and defaults change between releases.
Standards and references
- Alibaba Qwen: Qwen3.8-2.4T-A95B model card (accessed 2026-11-15)
- Alibaba Cloud: Model Studio pricing (accessed 2026-11-15)
- DeepSeek: DeepSeek-V4-Pro model card (accessed 2026-11-15)
- DeepSeek: Models and pricing (accessed 2026-11-15)
- Google: Gemma 4 model overview (accessed 2026-11-15)
- LG AI Research: K-EXAONE 2.0 model card (accessed 2026-11-15)
- Meta: Muse Spark Open model card (accessed 2026-11-15)
- MiniMax: MiniMax-M3 model card (accessed 2026-11-15)
- Mistral AI: Mistral Large 4 model card (accessed 2026-11-15)
- Mistral AI: Reasoning (accessed 2026-11-15)
- Moonshot AI: Kimi K3.1 Pro model card (accessed 2026-11-15)
- NVIDIA: Nemotron 3 Ultra model card (accessed 2026-11-15)
- OpenAI: gpt-oss-120b model page (accessed 2026-11-15)
- Tencent: Hy3 model card (accessed 2026-11-15)
- Z.ai: GLM-5.3 model card (accessed 2026-11-15)
- Z.ai: Pricing (accessed 2026-11-15)
Related NHI Mgmt Group resources: Open-Weight vs Closed Models Explained, Open-Weight Model Security Guide, Self-Hosted LLMs vs Hosted LLM APIs, Frontier AI Models Compared, Mixture of Experts Explained.