whatmodel.sh / Step 3
Step 3
↗ share
+ contribute
Step 3 is an open-weights language model from StepFun, released July 2025. Its published size is 321B parameters (38B active per token). Licence: apache-2.0.
Catalogue identity notes
Comparison-only. Arena entry 'step-3' (text, vision). Maker repo stepfun-ai/step3 @e2d67fc: licence tag apache-2.0. Parameters: StepFun: 321B total / 38B active (HF safetensors metadata 321.0B). Released 2025-07: month the maker's Hugging Face repo was created (createdAt 2025-07-28). Arena 'step-3' matches StepFun's API id step-3 (Step 3 reasoning VLM) and the open-weight release stepfun-ai/step3 ('Both the code repository and the model weights are released under the Apache License'). No longer on StepFun's API.
$ how it compares
The LLMs ranked either side of Step 3, out of 1132 with a score, then more from StepFun.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 300 |  TranslateGemma 4B | 59±10 | local | ||
| 301 |  DeepSeek V3 (Dec '24) | 59±1 | api $0.46/M +local | ||
| 302 | Step 3 | 59±3 | local | ||
| 303 |  FuseChat-Qwen-2.5-7B-Instruct | 59±2 | local | ||
| 304 |  Gemma 2 9B IT SimPO | 59±1 | local | ||
| more from StepFun | |||||
| 45 | Step 5 Preview | 75±1 | comparison only | ||
| 130 | Step 3.5 Flash (196B-A11B) | 68±1 | api $0.15/M +local | ||
| 151 | Step 3.7 Flash (198B-A11B) | 67±1 | api $0.44/M +local | ||
$ what it's good at
chat & writing▁▁▁▄▆▅▆█▆▆▆▂ #166 of 466 * · 57% of its score · 1 test
Beats MiniMax-M2, MiniCPM5 2B, Granite 4.2 30B here, though they rank higher overall. Trails LFM2.5 1.2B Thinking, LFM2.5 8B-A1B, LFM2.5 2.6B here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| LMArena (text) | 1349.4 ✎ | ▁▁▁▂▃▃▃▆▆▇█▃ beats 193/379 | 57% |
vision & documents▁▁▃▂▁▃▄▃▆▇█▄ #96 of 150 * · 40% of its score · 1 test
Beats DeepSeek V4 Flash Vision (Exp) here, though they rank higher overall. Trails GPT-4o (2024-11-20), Qwen3-VL 8B Thinking here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Arena vision (style controlled) | 1144.9 ✎ | ▁▁▂▂▂▂▄▅▅▆██ beats 44/145 | 40% |
$ where to run it
Other ways to run it
- Run it as published · needs ~327.59 GB, NVIDIA only · from huggingface.co/stepfun-ai
uv pip install -U vllm --torch-backend=auto --extra-index-url https://wheels.vllm.ai/nightly && vllm serve stepfun-ai/step3-fp8 --tensor-parallel-size 8 --reasoning-parser step3 --enable-auto-tool-choice --tool-call-parser step3 --gpu-memory-utilization 0.85 --max-num-batched-tokens 4096 --trust-remote-code
Weights and documentation: huggingface.co/stepfun-ai ↗
$ how the score was worked out · LMArena leaderboard dataset 97%
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
| source | results | share of the score |
|---|---|---|
 LMArena leaderboard dataset | 2 | 97% |
| prior · pull toward the middle | 3% | |
| weighted average | = 59.3 |
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json