This page answers one question: which downloadable-weight models are the strongest right now, and what are you allowed to do with them. It deliberately ignores whether they fit your graphics card. If hardware is the constraint, the VRAM tier index is the right page instead, and it answers a genuinely different question.
Every parameter count, licence and access flag below was read from the Hugging Face model API on 7 August 2026, not from memory. Parameter totals come from the safetensors.total field of each repository, licences from cardData.license or the repository’s own licence file, and the gated column from the gated field. Active-parameter figures for mixture-of-experts models come from the model cards, which is the only place most vendors state them.
The current field
| Model | Parameters (total / active) | Licence | Gated | Best at |
|---|---|---|---|---|
moonshotai/Kimi-K3 | 2.78T / 104B | Kimi K3 Licence (custom) | No | Long-horizon agentic coding, native multimodal |
deepseek-ai/DeepSeek-V4-Pro | 1.60T / 49B | MIT | No | Reasoning and knowledge work at 1M context |
zai-org/GLM-5.2 | 753B | MIT | No | The strongest fully permissive frontier model |
MiniMaxAI/MiniMax-M3 | 427B / ~23B | MiniMax Community (custom) | No | Text, image and video understanding at 1M context |
Qwen/Qwen3.5-397B-A17B | 403B / 17B | Apache-2.0 | No | The largest Apache-licensed Qwen |
meta-llama/Llama-4-Maverick-17B-128E-Instruct | 402B / 17B | Llama 4 Community | Yes (manual) | Multimodal chat, if you accept the terms |
deepseek-ai/DeepSeek-V4-Flash-0731 | 304B | MIT | No | Frontier-adjacent output at low activation cost |
thinkingmachines/Inkling-Small | 266B | Apache-2.0 | No | Text, image and audio input under a plain licence |
mistralai/Mistral-Medium-3.5-128B | 128B (dense) | Modified MIT, revenue-capped | No | Dense multimodal reasoning at 256k context |
nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 | 124B / 12B | NVIDIA Nemotron Open Model Licence | No | Throughput, via interleaved Mamba-2 and MoE layers |
mistralai/Mistral-Small-4-119B-2603 | 119B / 6.5B | Apache-2.0 | No | Permissive multimodal at low activation cost |
poolside/Laguna-S-2.1 | 118B / 8B | OpenMDW-1.1 | No | Agentic coding |
openai/gpt-oss-120b | 117B / 5.1B | Apache-2.0 | No | Strong general model that fits one 80GB accelerator |
Qwen/Qwen3-Coder-Next | 80B / 3B | Apache-2.0 | No | Cheap agentic coding |
google/gemma-4-31B-it | 31.3B | Apache-2.0 | No | Ungated permissive multimodal at a workstation size |
Qwen/Qwen3.6-27B | 27.8B (dense) | Apache-2.0 | No | The strongest dense model under 30B |
Totals are rounded from the byte-exact safetensors.total values. Dense models have no separate active count, so only one figure is given.
The trillion-parameter tier is real and you cannot run it
The three most capable open-weight models on the list are physically out of reach of any consumer card, and pretending otherwise does readers no favours.
Kimi K3 reports 2,779,931,837,184 parameters in safetensors.total, with the model card putting activated parameters at 104B and describing a Stable LatentMoE design that selects 16 of 896 experts per token. The weights ship quantization-aware-trained at MXFP4, which is the only reason the download is as small as it is. DeepSeek-V4-Pro is 1.60T total with 49B activated per its card. Moonshot’s Kimi-K2.7-Code sits at 1.03T and inclusionAI/Ring-2.5-1T at 1.01T, both verified the same way.
Nothing in that group runs on a 32GB card, or on eight of them. These are multi-node models. Their practical availability to most readers is through a hosted API, which makes “open weights” a licensing and auditability property rather than a self-hosting one. That is still worth something: the weights can be inspected, fine-tuned by anyone with the hardware, and served by a provider other than the lab that made them. It is not the same thing as running it yourself.
The middle of the table is where self-hosting becomes plausible for an organisation rather than an individual. openai/gpt-oss-120b is explicit about this in its card, describing itself as targeting a single 80GB accelerator, and its 5.1B active parameters keep the compute cost low even though the weights are 117B.
Licence is the part people get wrong
“Open weights” and “open source” are not the same claim, and the gap is widest exactly where the models are strongest.
Four of the sixteen carry MIT and six carry Apache-2.0, which are the straightforward cases. GLM-5.2’s card states its position bluntly, describing an MIT licence with no regional limits. DeepSeek’s V4 line is MIT across DeepSeek-V4-Pro, DeepSeek-V4-Flash and the newer DeepSeek-V4-Flash-0731. Qwen’s frontier and small models alike return apache-2.0 with a license_link pointing at a LICENSE file inside the repository.
The custom licences need reading:
- Kimi K3 ships under a bespoke “Kimi K3 Licence”. Its clause 2 requires a separate agreement with Moonshot AI if the licensee operates a model-as-a-service business with more than 20 million US dollars of aggregate revenue over any consecutive 12 months, and clause 3 requires prominent “Kimi K3” attribution in products above 100 million monthly active users or 20 million dollars of monthly revenue. Internal use is carved out.
- MiniMax M3 returns
license_name: minimax-community. The licence file grants rights “for non-commercial purposes” by default, requires a “Built with MiniMax M3” notice for commercial use, and requires prior written authorisation above 20 million dollars of yearly revenue. - Mistral Medium 3.5 128B returns a bare
otherin the API. The repository’s licence file is a Modified MIT that withdraws all rights if the global consolidated monthly revenue of your company exceeds 20 million dollars for the preceding month. Its siblingMistral-Small-4-119B-2603is plain Apache-2.0, so two models from the same vendor sit on opposite sides of that line. - NVIDIA Nemotron 3 Super returns
license_name: nvidia-nemotron-open-model-license, a vendor agreement hosted on nvidia.com rather than an OSI-approved licence. - Laguna S 2.1 returns
openmdw-1.1. Its licence file is the OpenMDW License Agreement version 1.1, which grants permission to deal in the “Model Materials” without restriction across copyright, patent, database and trade-secret rights.
Only one repository in the table is access-gated. meta-llama/Llama-4-Maverick-17B-128E-Instruct returns gated: "manual", meaning a human at Meta approves each request, and its licence is the Llama 4 Community agreement rather than an open-source licence. meta-llama/Llama-4-Scout-17B-16E-Instruct is the same on both counts.
Gating is not predictable from the vendor. Google’s gemma-4-31B-it returns license: apache-2.0 and gated: false, while the older google/gemma-3n-E4B-it in the same account returns license: gemma and gated: "manual". Two Google repositories, opposite answers. Check the specific repository rather than the brand. One wrinkle on Gemma 4: the card declares Apache-2.0 while also carrying a license_link to a Gemma 4 licence page on ai.google.dev, and there is no LICENSE file in the repository itself, so the API field is the only in-repo statement available.
Benchmarks, and why there are so few here
Almost every comparison table published with these models is produced by the lab that trained one of the entries. Two were read directly today and are labelled as such.
Moonshot’s own Kimi K3 card reports Terminal-Bench 2.1 at 88.3 for K3, against 88.8 for GPT-5.6 Sol, 88.0 for Claude Fable 5, 84.6 for Claude Opus 4.8 and 82.7 for GLM-5.2, with a footnote conceding that K3 was run in Moonshot’s own Kimi Code harness while rivals were scored in theirs. Z.ai’s GLM-5.2 card reports SWE-bench Pro at 62.1 for GLM-5.2, 59 for MiniMax M3 and 55.4 for DeepSeek-V4-Pro. Both are vendor self-reported, both grade the publisher favourably, and no independent re-run of either is cited here because none was re-sourced today. Treat the ordering as a hint and the margins as marketing.
What was searched for and not found
Qwen3.8-Max was announced on 3 August with open weights promised, and the site covered it. As of 7 August the weights are not on Hugging Face: Qwen/Qwen3.8-Max, Qwen/Qwen3.8-27B and Qwen/Qwen3.8 all return no repository, and the Qwen account sorted by creation date shows nothing newer than late June. Third-party accounts host models named “Qwen3.8” distills, but the official release has not landed. The largest current Qwen open weights remain Qwen3.5-397B-A17B.
Meta has published nothing newer than the April 2025 Llama 4 repositories in the meta-llama account. A search for meta-llama/Llama-5 returns no repository. xAI’s xai-org account tops out at grok-2, uploaded in August 2025.
Where to go from here
If the constraint is your hardware rather than the state of the art, start from the VRAM tier index, which maps card memory to a specific model and quantization for each task. If the question is how much quality a given precision costs, the quantization comparison covers GGUF, AWQ, GPTQ and MLX and the real bits per weight behind each name.
If the constraint is legal rather than technical, the shortest safe path through the table is the MIT and Apache column: GLM-5.2 and the DeepSeek V4 line at the top, Qwen3.5-397B-A17B, Inkling-Small, Mistral-Small-4, gpt-oss-120b and Qwen3-Coder-Next in the middle, Gemma 4 31B and Qwen3.6-27B at the bottom. Everything else on the list carries a condition that a lawyer should read before it reaches production.
Verified against the Hugging Face model API on 7 August 2026. Repositories change; re-check the licence field before shipping.