The open weight model scene, mid-2025: who's winning?
Meta, Alibaba, Mistral, DeepSeek, and 12 others are all releasing open weight models. I ranked them by Chatbot Arena Elo, Hugging Face downloads, and community adoption. Llama still leads downloads, but Qwen is closing fast.
The open weight model space is crowded. Sixteen organizations are actively releasing models. I ranked them across three dimensions to figure out who's actually winning.
The contenders
| Organization | Country | Latest model | Total params | Open weight? | |-------------|---------|-------------|-------------|-------------| | Meta AI | US | Llama 4 Maverick | 400B (17B active) | Yes | | Alibaba/Qwen | China | Qwen3 235B | 235B (22B active) | Yes | | DeepSeek | China | DeepSeek R1 | 671B (37B active) | Yes | | Mistral AI | France | Mistral Large 3 | ~70B | Partial | | MiniMax | China | MiniMax-01 | 456B (45.9B active) | Yes | | Microsoft | US | Phi-4 | 14B | Yes | | Google | US | Gemma 2 | 27B | Yes | | Cohere | Canada | Command R+ | 104B | Yes | | 01.AI | China | Yi-Lightning | Unknown | No (API only) | | Stability AI | UK | Stable LM 2 | 12B | Yes |
Ranking 1: Chatbot Arena Elo (quality)
| Rank | Model | Elo | |------|-------|-----| | 1 | DeepSeek R1 | 1255 | | 2 | Qwen3 235B | 1256 | | 3 | DeepSeek V3 | 1258 | | 4 | Llama 4 Maverick | 1248 | | 5 | MiniMax-01 | 1242 | | 6 | Mistral Large 3 | 1234 | | 7 | Llama 3.1 405B | 1230 | | 8 | Qwen 2.5 72B | 1220 |
Sources: LMSYS Chatbot Arena, June 2025.
Chinese models hold 3 of the top 5 open weight positions. DeepSeek and Qwen dominate the quality ranking. Meta's Llama 4 Maverick is 4th, which is respectable but no longer the clear leader.
Ranking 2: Hugging Face downloads (adoption)
| Rank | Model family | Monthly downloads (Jun 2025) | |------|-------------|---------------------------| | 1 | Llama (all versions) | 142M | | 2 | Qwen (all versions) | 98M | | 3 | Mistral (all versions) | 61M | | 4 | DeepSeek (all versions) | 54M | | 5 | Phi (all versions) | 41M | | 6 | Gemma (all versions) | 28M | | 7 | MiniMax | 12M |
Sources: Hugging Face download counters, June 2025 snapshot.
Llama still leads downloads by a wide margin. 142M vs Qwen's 98M. But the gap is narrowing. In January 2025, Llama had 2.8x more downloads than Qwen. Now it's 1.45x.
Mistral at 61M holds steady. DeepSeek at 54M surged after R1's launch. Phi at 41M is impressive for a small model family. Microsoft's strategy of targeting the under 14B space is paying off in adoption.
Ranking 3: Community adoption
| Model family | Fine-tuned variants on HF | Active repos on GitHub | Community engagement | |-------------|--------------------------|----------------------|---------------------| | Llama | 8,400+ | 2,100+ | Largest | | Qwen | 3,200+ | 800+ | Growing fast | | Mistral | 2,800+ | 600+ | Stable | | DeepSeek | 1,400+ | 400+ | Surging | | Phi | 900+ | 300+ | Moderate |
Sources: Hugging Face model hub search, GitHub search, June 2025.
Llama's community footprint is still 2-3x larger than any competitor. 8,400 fine-tuned variants is a massive moat. When developers build on Llama, they create tutorials, tools, and integrations that make the next developer more likely to also choose Llama.
Qwen is growing fastest in absolute terms, driven partly by strong Chinese developer adoption and partly by genuine quality improvements.
Combined ranking
| Overall rank | Organization | Quality | Adoption | Community | Trend | |-------------|-------------|---------|----------|-----------|-------| | 1 | Meta (Llama) | 4th | 1st | 1st | Stable | | 2 | Alibaba (Qwen) | 2nd | 2nd | 2nd | Rising | | 3 | DeepSeek | 1st | 4th | 4th | Rising fast | | 4 | Mistral AI | 6th | 3rd | 3rd | Stable | | 5 | Microsoft (Phi) | N/A (small models) | 5th | 5th | Stable |
Meta leads overall because community size and adoption count for more than raw benchmarks in practice. But the quality crown now belongs to DeepSeek and Qwen.
The tension between "best quality" and "largest adoption" is real. Llama is the jQuery of LLMs: not the absolute best at anything, but everyone knows it and has tools for it.
My prediction for end of 2025
| Prediction | |-----------| | Llama retains #1 downloads but Qwen closes to within 1.2x | | DeepSeek R2 takes #1 on open model quality | | Mistral consolidates in the enterprise niche | | A non-Chinese, non-Meta lab breaks into the top 5 (possibly Cohere or a new entrant) | | Total open weight model downloads exceed 500M per month |
The open weight space is the most competitive part of AI right now. Sixteen organizations, six countries, hundreds of billions in total investment. It's beautiful chaos.
My "open model leaderboard" spreadsheet has 47 rows. A year ago it had 12.
If you found this interesting, you might also like:
- I counted every AI startup that raised money in Q1 2021. The numbers are strange.
- AI funding in Q1 2023 is absolutely bonkers. Let me show you the numbers.
- The AI chip market in 2024: not just NVIDIA anymore
- The GPT-3 API waitlist is 6 months long. Here's what the early data looks like.
- Hugging Face just hit 10,000 models. Here's what the model zoo looks like.
-- dataku