Open-weight models you can host yourself: with licence, size, context length and the memory the weights need. The facts come straight from Hugging Face and are re-checked every two weeks.
Last updated: · Next update: · Source: Hugging Face API
| Model | Parameters | Type | Context | Licence | HLE | On HF since |
|---|---|---|---|---|---|---|
| DeepSeek V4.1 Flash DeepSeek | 763.2B | Mixture of Experts | 1M | MIT | – | Sept 2026 |
| GLM-5.3 Z.ai (Zhipu) | 753.3B | Mixture of Experts | 1M | Custom licence (GLM-5.3) | 39.1 % | Aug 2026 |
| GLM-5.3 Flash Z.ai (Zhipu) | 321.3B | Mixture of Experts | 1M | MIT | – | Aug 2026 |
| DeepSeek V4 Pro DeepSeek | 1.7T | Mixture of Experts | 1M | MIT | 32.4 % | Aug 2026 |
| Qwen3.8 2.4T-A95B Alibaba (Qwen) | 2.4Tof which active 95B | Mixture of Experts | 256K | Custom licence (Qwen3.8-Max) | – | Aug 2026 |
| Qwen3.8 27B Alibaba (Qwen) | 27.8B | Dense | 256K | Apache 2.0 | – | Aug 2026 |
| Kimi K3 Moonshot AI | 2.8T | Mixture of Experts | 1M | Custom licence (Kimi K3) | 41.1 % | Jun 2026 |
| Gemma 4 12B Google | 12B | Dense | 256K | Apache 2.0 | – | May 2026 |
| Qwen3.6 35B-A3B Alibaba (Qwen) | 36Bof which active 3B | Mixture of Experts | 256K | Apache 2.0 | – | Apr 2026 |
| Mistral Medium 3.5 Mistral AI | 127.7B | Dense | 256K | Custom licence (see model card) | – | Mar 2026 |
| Gemma 4 31B Google | 31.3B | Dense | 256K | Apache 2.0 | – | Mar 2026 |
| Gemma 4 26B-A4B Google | 25.8Bof which active 4B | Mixture of Experts | 256K | Apache 2.0 | – | Mar 2026 |
| Gemma 4 E4B Google | 8B | Dense | 128K | Apache 2.0 | – | Mar 2026 |
| Mistral Small 4 Mistral AI | 119.4B | Mixture of Experts | 1M | Apache 2.0 | – | Jan 2026 |
| Mistral Large 3 Mistral AI | 675B | Mixture of Experts | – | Apache 2.0 | – | Nov 2025 |
| gpt-oss-120b OpenAI | 116.8B | Mixture of Experts | 128K | Apache 2.0 | – | Aug 2025 |
| gpt-oss-20b OpenAI | 20.9B | Mixture of Experts | 128K | Apache 2.0 | – | Aug 2025 |
| Llama 4 Scout Meta | 108.6Bof which active 17B | Mixture of Experts | 10M | Llama 4 Community License | – | Apr 2025 |
| Llama 4 Maverick Meta | 401.6Bof which active 17B | Mixture of Experts | – | Llama 4 Community License | – | Apr 2025 |
HLE values come from the Statistics.
The table shows how much memory the weights alone need (parameters × bytes per parameter). Add the context cache (KV cache) and runtime overhead: plan for at least 20% headroom.
| Model | 16-bit (BF16) | 8-bit | 4-bit |
|---|---|---|---|
| Gemma 4 E4B | 16 GB | 8 GB | 4 GB |
| Gemma 4 12B | 24 GB | 12 GB | 6 GB |
| gpt-oss-20b | 41.8 GB | 20.9 GB | 10.4 GB |
| Gemma 4 26B-A4B | 51.6 GB | 25.8 GB | 12.9 GB |
| Qwen3.8 27B | 55.6 GB | 27.8 GB | 13.9 GB |
| Gemma 4 31B | 62.6 GB | 31.3 GB | 15.7 GB |
| Qwen3.6 35B-A3B | 72 GB | 36 GB | 18 GB |
| Llama 4 Scout | 217 GB | 109 GB | 54.3 GB |
| gpt-oss-120b | 234 GB | 117 GB | 58.4 GB |
| Mistral Small 4 | 239 GB | 119 GB | 59.7 GB |
| Mistral Medium 3.5 | 255 GB | 128 GB | 63.9 GB |
| GLM-5.3 Flash | 643 GB | 321 GB | 161 GB |
| Llama 4 Maverick | 803 GB | 402 GB | 201 GB |
| Mistral Large 3 | 1.4 TB | 675 GB | 338 GB |
| GLM-5.3 | 1.5 TB | 753 GB | 377 GB |
| DeepSeek V4.1 Flash | 1.5 TB | 763 GB | 382 GB |
| DeepSeek V4 Pro | 3.3 TB | 1.7 TB | 825 GB |
| Qwen3.8 2.4T-A95B | 4.9 TB | 2.4 TB | 1.2 TB |
| Kimi K3 | 5.6 TB | 2.8 TB | 1.4 TB |
Apache 2.0 and MIT allow commercial use without asking. Custom licences (Qwen3.8-Max, GLM-5.3, Kimi K3, Mistral Medium, Llama 4) carry special terms that should be read before a product launch.
DeepSeek · MIT · 763.2B · 1M
DeepSeek's current Flash model with a Mixture-of-Experts architecture and 1M tokens of context. The API supports tool calls, JSON output, image input and a thinking mode.
Model card on Hugging Face · Downloads: 1,355,278
Z.ai (Zhipu) · Custom licence (GLM-5.3) · 753.3B · 1M
Z.ai's flagship with open weights (about 753B parameters, 1M tokens of context). Licence: GLM-5.3 licence, not MIT.
Model card on Hugging Face · Downloads: 1,670,161
Z.ai (Zhipu) · MIT · 321.3B · 1M
Smaller variant of GLM-5.3 under the MIT licence (about 321B parameters, 1M tokens of context).
Model card on Hugging Face · Downloads: 6,479,110
DeepSeek · MIT · 1.7T · 1M
Large Pro model of the V4 family (version 0813) with 1M tokens of context. The API documentation lists tool calls and JSON output, but no image input.
Model card on Hugging Face · Downloads: 87,569
Alibaba (Qwen) · Custom licence (Qwen3.8-Max) · 2.4T · 256K
With about 2.4 trillion parameters (about 95B active according to its name) the largest Qwen model with open weights. The licence is not a standard one: read the licence text before commercial use.
Model card on Hugging Face · Downloads: 32,082
Alibaba (Qwen) · Apache 2.0 · 27.8B · 256K
Dense model with just under 28B parameters and 262K tokens of context under Apache 2.0. Fits on a single 24 GB GPU in 4-bit quantisation (weights alone about 14 GB).
Model card on Hugging Face · Downloads: 6,768,654
Moonshot AI · Custom licence (Kimi K3) · 2.8T · 1M
Open-weight model from Moonshot AI with about 2.8 trillion parameters and 1M tokens of context. Realistically only for data-centre deployment; check the licence before commercial use.
Model card on Hugging Face · Downloads: 1,160,630
Google · Apache 2.0 · 12B · 256K
Mid-sized Gemma 4 for laptops and small servers (262K tokens of context). Weights alone need about 6 GB in 4-bit.
Model card on Hugging Face · Downloads: 1,603,246
Alibaba (Qwen) · Apache 2.0 · 36B · 256K
Mixture-of-Experts model with about 36B parameters, of which about 3B are active according to its name. Inference is fast, but memory demand follows all weights.
Model card on Hugging Face · Downloads: 3,545,718
Mistral AI · Custom licence (see model card) · 127.7B · 256K
Dense model with about 128B parameters and 262K tokens of context. The weights are open, but the licence is not a standard one like Apache 2.0.
Model card on Hugging Face · Downloads: 72,473
Google · Apache 2.0 · 31.3B · 256K
Largest dense model of the Gemma 4 family with 262K tokens of context. Google replaced the earlier custom licence with Apache 2.0; the model processes images and video.
Model card on Hugging Face · Downloads: 9,619,659
Google · Apache 2.0 · 25.8B · 256K
Mixture-of-Experts variant of Gemma 4 with about 26B parameters, about 4B of them active according to its name, 262K tokens of context.
Model card on Hugging Face · Downloads: 11,854,067
Google · Apache 2.0 · 8B · 128K
Small Gemma 4 model (4.5B effective parameters) for phones and edge devices; processes audio in addition to image and video.
Model card on Hugging Face · Downloads: 4,305,979
Mistral AI · Apache 2.0 · 119.4B · 1M
Mixture-of-Experts model with about 119B parameters and 1M tokens of context under Apache 2.0. Despite the name "Small", the weights need several GPUs.
Model card on Hugging Face · Downloads: 79,509
Mistral AI · Apache 2.0 · 675B · –
At 675B parameters, Mistral's large open-weight model (November 2025) under Apache 2.0.
Model card on Hugging Face · Downloads: 2,037
OpenAI · Apache 2.0 · 116.8B · 128K
Open model from OpenAI (August 2025) with about 117B parameters and 131K tokens of context under Apache 2.0. Not updated since its release, so older than the other models.
Model card on Hugging Face · Downloads: 3,845,925
OpenAI · Apache 2.0 · 20.9B · 128K
Small sibling model with about 21B parameters and 131K tokens of context under Apache 2.0 (August 2025).
Model card on Hugging Face · Downloads: 5,950,578
Meta · Llama 4 Community License · 108.6B · 10M
Smaller Llama 4 model with about 109B parameters, 17B of them active (April 2025). Meta quoted up to 10M tokens of context. Access requires approval.
Gated: access after approval by Meta.
Model card on Hugging Face · Downloads: 163,461
Meta · Llama 4 Community License · 401.6B · –
Mixture-of-Experts model with about 400B parameters, 17B of them active (April 2025). Meta has published no newer Llama release on Hugging Face (as of 10 Oct 2026). Access requires approval.
Gated: access after approval by Meta.
Model card on Hugging Face · Downloads: 9,467
With 4-bit quantisation the weights need about half a byte per parameter. With roughly 19 GB for the weights (about 80% of 24 GB, the rest is for the context cache) these fit: Qwen3.6 35B-A3B (36B), Gemma 4 31B (31.3B), Qwen3.8 27B (27.8B), Gemma 4 26B-A4B (25.8B), gpt-oss-20b (20.9B), Gemma 4 12B (12B), Gemma 4 E4B (8B). Larger models need several cards or offloading.
The trained weights are available for download and can be run yourself. That is not automatically open source: training data and code are often not open, and the licence decides about commercial use.
Models under Apache 2.0 or MIT allow commercial use without asking: DeepSeek V4.1 Flash, DeepSeek V4 Pro, Qwen3.8 27B, Qwen3.6 35B-A3B, Gemma 4 31B, Gemma 4 26B-A4B, Gemma 4 12B, Gemma 4 E4B, Mistral Small 4, Mistral Large 3, GLM-5.3 Flash, gpt-oss-120b, gpt-oss-20b. For models with their own licence, read the licence itself before using them in a product.
Licence, parameter count, context length and dates are re-read from Hugging Face every two weeks. Last update: 10 October 2026.