whatmodel.sh / Step 3.7 Flash (198B-A11B)
Step 3.7 Flash (198B-A11B)
↗ share
+ contribute
Step 3.7 Flash (198B-A11B) is an open-weights language model from StepFun, released May 2026. Its published size is 198B parameters (11B active per token) and a 256k-token context window. Licence: apache-2.0.
Catalogue identity notes
Comparison-only. Official model card (huggingface.co/stepfun-ai/Step-3.7-Flash @ 5f6244077ac62e04eec3f320501ff8c2b293373a, created 2026-05-23, Apache-2.0): 198B-parameter sparse MoE vision-language model (196B language backbone + 1.8B vision encoder), about 11B active per token, reasoning levels low/medium/high, 256k context (context not recorded here: no approved context source). MathArena config stepfun/3.7-flash: API model 'step-3.7-flash', reasoning_effort high, date 2026-05-29. Distinct from step-3-5-flash.
$ how it compares
The LLMs ranked either side of Step 3.7 Flash (198B-A11B), out of 828 with a score, then more from StepFun.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 130 |  Laguna M.1 | 67±2 | api price unknown +local | ||
| 131 |  Qwen3.7 Flash | 67±1 | api $0.055/M | ||
| 132 | Step 3.7 Flash (198B-A11B) | 67±1 | api $0.44/M +local | ||
| 133 |  Gemma 4 31B | 67±1 | api $0.19/M +local | ||
| 134 |  Gemini 3.1 Flash-Lite | 67±1 | api $0.56/M | ||
| more from StepFun | |||||
| 45 | Step 5 Preview | 75±1 | comparison only | ||
| 115 | Step 3.5 Flash (196B-A11B) | 68±1 | api $0.15/M +local | ||
$ what it's good at
agent tasksnot ranked on this task yet · 3 tests
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Terminal-Bench 2.1 lab | 59.5% ✎ | ▁·▁·▁▁▄▃▄█▃▇ beats 23/56 | — |
| SWE-bench Pro (split unspecified) lab | 56.3% ✎ | ▃·▁▁▁▂▂▃█▇▃▁ beats 33/55 | — |
| Toolathlon (Tool Decathlon) lab | 49.5% ✎ | ▂▂▄▂▄▄▆▂█▆██ beats 16/27 | — |
codingnot ranked on this task yet · 1 test
| test | result | vs the others in its run | counts |
|---|---|---|---|
| SWE-bench Pro (split unspecified) lab | 56.3% ✎ | ▃·▁▁▁▂▂▃█▇▃▁ beats 33/55 | — |
math▁▂▃▅▆▅▄█▄▁▁▁ #102 of 398 · 93% of its score · 11 tests
Beats Claude Sonnet 4.6, MiniMax-M3, Nemotron 3 Ultra 550B-A55B here, though they rank higher overall. Trails Gemini 2.5 Flash, GLM-4.5-Air, Granite 4.2 30B here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| MathArena ArXivMath June 2026 (48 problems) | 30.6% ✎ | ▂···▄·██▆▆▄▂ beats 2/19 | 25% |
| MathArena ArXivMath April 2026 (40 problems) | 35.8% ✎ | ▁·····▅██▄·▁ beats 3/20 | 20% |
| MathArena ArXivMath February 2026 (32 problems) | 32.0% ✎ | ▁▁▁▁█▇▄▃▁▃·▃ beats 7/24 | 11% |
| MathArena ArXivMath March 2026 (30 problems) | 45.0% ✎ | █▃▅·█▃▃·▅·▃▃ beats 1/14 | 11% |
| MathArena ArXivMath May 2026 (40 problems) | 46.7% ✎ | ▁··▁▄█▄▁·▃·▁ beats 3/17 | 7.9% |
| MathArena BrokenArXiv April 2026 (61 problems) | 18.9% ✎ | ▂▂▆█▃▃▂▂▂··▂ beats 5/18 | 3.9% |
| MathArena BrokenArXiv June 2026 (54 problems) | 15.7% ✎ | ▄··█▂▂▂▆▄··▂ beats 5/14 | 3.2% |
| MathArena BrokenArXiv May 2026 (50 problems) | 8.5% ✎ | ██·▄▄▄····▄▄ beats 0/8 | 3.1% |
3 more tests
| test | result | vs the others in its run | counts |
|---|---|---|---|
| MathArena ArXivLean June 2026 (46 problems) | 0.0% ✎ | ▄····▄▄███·▄ beats 0/9 | 2.7% |
| MathArena BrokenArXiv March 2026 (56 problems) | 10.7% ✎ | █·▆█▂·▂▂···▂ beats 4/14 | 2.4% |
| MathArena BrokenArXiv February 2026 (31 problems) | 12.1% ✎ | ▃▂▂▃█▃▂·▃··▂ beats 9/16 | 2.3% |
other tests5 not in a task · 6.7% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| WideSearch corrected dataset | 56.8% ✎ | ▃······▃▃▃█▃ beats 1/7 | 3.5% |
| SkillsBench 87-task subset | 24.9% ✎ | ▄··▄▄▄···█·█ beats 2/7 | 3.2% |
| SimpleVQA with tools (search) lab | 79.2% ✎ | no other model yet | — |
| V* Bench with Python tool lab | 95.3% ✎ | no other model yet | — |
| ClawEval 1.1 (metric unspecified) lab | 67.1% ✎ | no other model yet | — |
$ where to run it
Through an API
api · $0.44/M · $0.2 in / $1.15 out per 1M tokens · 256k context · platform.stepfun.ai ↗ · price sourceUSD per 1M tokens, StepFun Open Platform (international), step-3.7-flash (Multimodal Reasoning Models table): input on cache miss $0.20, output $1.15 (output includes reasoning). Cache-hit input ($0.04/1M) excluded. Currency: USD. StepFun official pricing page, observed 2026-09-26.
llama.cpp · recommended
version
- Install llama.cpp · Homebrew and winget installs name the commands llama-cli and llama-server instead.macOS / Linux
curl -LsSf https://llama.app/install.sh | shWindowsirm https://llama.app/install.ps1 | iex - Run Step 3.7 Flash (198B-A11B) · needs ~115.47 GB · from huggingface.co/stepfun-aichat in the terminal
llama cli -hf stepfun-ai/Step-3.7-Flash-GGUF:Q4_K_S - Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf stepfun-ai/Step-3.7-Flash-GGUF:Q4_K_S
LM Studio
version
- Install LM Studio · Open the app once before using the lms command.
- Run Step 3.7 Flash (198B-A11B) · needs ~115.47 GB · from huggingface.co/stepfun-ai
- Or download it in the terminal
lms get https://huggingface.co/stepfun-ai/Step-3.7-Flash-GGUF@Q4_K_S -y - Then run lms ls to see its name, and lms chat followed by that name
Weights and documentation: huggingface.co/stepfun-ai ↗
$ how the score was worked out · MathArena ArXivMath / BrokenA… 93%, inclusionAI/Ling-3.0-flash mo… 7%
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 |
|---|---|---|
 MathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasets | 11 | 93% |
| inclusionAI/Ling-3.0-flash model card | 2 | 7% |
| prior · pull toward the middle | 0% | |
| weighted average | = 67.0 |
not counted (6)
| stepfun-ai/Step-3.7-Flash model card | SWE-bench Pro (split unspecified) [correct] | 56.3% | fewer than three models have this test yet, or this source ran it for this model alone |
| stepfun-ai/Step-3.7-Flash model card | Terminal-Bench 2.1 [correct] | 59.5% | fewer than three models have this test yet, or this source ran it for this model alone |
| stepfun-ai/Step-3.7-Flash model card | Toolathlon (Tool Decathlon) [correct] | 49.5% | fewer than three models have this test yet, or this source ran it for this model alone |
| stepfun-ai/Step-3.7-Flash model card | SimpleVQA with tools (search) [correct] | 79.2% | fewer than three models have this test yet, or this source ran it for this model alone |
| stepfun-ai/Step-3.7-Flash model card | V* Bench with Python tool [correct] | 95.3% | fewer than three models have this test yet, or this source ran it for this model alone |
| stepfun-ai/Step-3.7-Flash model card | ClawEval 1.1 (metric unspecified) [correct] | 67.1% | fewer than three models have this test yet, or this source ran it for this model alone |
$ 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