Open-Weight Models
The models you can download, self-host, and fine-tune. As of July 2026. Two things to know: (1) "open weight" ≠ open source. Some ship under permissive licenses (Apache-2.0 / MIT), others under custom, use-restricted ones, so the License column matters. Check it before you build on a model. (2) The open frontier is now led by Chinese labs (DeepSeek, Qwen, Kimi, GLM), while Meta has pivoted toward closed models.
| Model (latest open) | Released | Params (total / active) | License | Context | Best at | Run with |
|---|---|---|---|---|---|---|
| DeepSeek V4 (Pro / Flash) | Apr 2026 | Pro 1.6T/49B · Flash 284B/13B | MIT ✅ | 1M | Reasoning, coding, cheap inference | vLLM, SGLang |
| Qwen3.6 (35B-A3B) | Apr 2026 | 35B / 3B active | Apache-2.0 ✅ | 262K (~1M YaRN) | Agentic coding, repo-level reasoning | Ollama, vLLM, SGLang |
| Kimi K3 (Moonshot) | Jul 2026 | 2.8T / ~50B active | Modified MIT ✅ | 1M | Coding/agentic, near-Opus-4.8 | vLLM, SGLang |
| GLM-5.2 (Zhipu / Z.ai) | 2026 | ~744B / 40B active | MIT ✅ | ~1M | Coding/agentic/reasoning | vLLM, SGLang |
| Llama 4 Scout (Meta) | Apr 2025 | 109B / 17B active | Llama 4 Community ⚠️ | 10M | Very long context, single-GPU | Ollama, vLLM |
| Llama 4 Maverick (Meta) | Apr 2025 | 400B / 17B active | Llama 4 Community ⚠️ | 1M | Multimodal, general chat | vLLM |
| Mistral Large 3 | Dec 2025 | 675B / 41B active | Apache-2.0 ✅ | Large | Multimodal, multilingual | vLLM |
| Ministral 3 (3/8/14B) | Dec 2025 | dense | Apache-2.0 ✅ | n/a | Best small cost/perf, edge | Ollama, llama.cpp |
| Gemma 4 (Google) | Apr 2026 | 12B (+ MoE 26B-A4B) | Apache-2.0 ✅ | 256K | On-device → server, multimodal | Ollama, llama.cpp, MLX |
| Qwen3 (0.6B-235B) | 2025 | dense + MoE | Apache-2.0 ✅ | 32K-128K+ | Widest size range, 119 languages | Ollama, vLLM |
| Phi-4 family (Microsoft) | 2024-26 | 3.8B-~15B | MIT ✅ | 16K-128K | Reasoning-per-parameter, on-device | Ollama, llama.cpp |
✅ permissive · ⚠️ source-available with use restrictions (check the model card).
Rough VRAM rule of thumb: a dense model needs ~(params × bytes-per-param) of VRAM, e.g. a 27B model at 4-bit quantization ≈ 16 GB. MoE models load all experts into memory but only compute the active set, so they're fast but still memory-hungry.
Sources: Qwen · DeepSeek · Mistral · Gemma · Kimi · GLM · Llama.
Track open models live: Hugging Face trending (the canonical release feed; the original Open LLM Leaderboard was archived in 2025), LMArena (includes open models), llm-stats.com.