05 / OBSERVATORY · MODELS Back to home
Open models: the 2026 map.
In 2026 open-weight models with MIT or Apache licenses cover every size, from the phone to the data center. Here are the ones we use as a reference, grouped by the hardware they run on.
Updated September 2026
Updated September 2026
Server GPU models. Mixture-of-experts.
Total / active parameters, license, weight memory, release.
| Model | Total / active params | License | Weight memory | Release |
|---|---|---|---|---|
| DeepSeek V4-Pro | 1,6 T / 49 B | MIT | ~865 GB | Aug 2026 |
| DeepSeek V4-Flash | 284 B / 13 B | MIT | ~170 GB | Jul 2026 |
| GLM-5.2 (Z.ai) | ~750 B / 40 B | MIT | ~400 GB in FP4 | Jun 2026 |
| Qwen3.5-397B (Alibaba) | 397 B / 17 B | Apache 2.0 | ~220 GB in Q4 | Feb 2026 |
| Mistral Large 3 | 675 B / 41 B | Apache 2.0 | ~350 GB in FP4 | Dec 2025 |
| gpt-oss-120b (OpenAI) | 117 B / 5 B | Apache 2.0 | ~80 GB | Aug 2025 |
Edge, portables, phones.
Sizes, license, strengths.
| Model | Sizes | License | Strengths |
|---|---|---|---|
| Gemma 4 (Google) | E2B, E4B, 26B MoE, 31B | Apache 2.0 | images, video and audio; runs offline even on a Raspberry Pi |
| Qwen3.5 small | 0,8–9 B | Apache 2.0 | 201 languages, 262K context |
| Ministral 3 | 3, 8, 14 B | Apache 2.0 | European, reads images |
| Granite 4.0 Nano (IBM) | 350 M, 1 B | Apache 2.0 | runs in the browser |
| SmolLM3 (Hugging Face) | 3 B | Apache 2.0 | fully open |
| LFM2.5 (Liquid AI) | 0,2–8 B | licenza propria: gratis sotto 10 M$ di fatturato | very fast on device |