[whatmodel.sh](../../README.md) / Step 3.5 Flash (196B-A11B) # Step 3.5 Flash (196B-A11B) StepFun · released 2026-02 · 196B params (11B active) · apache-2.0 **score 67.1** ±0.9 · #113 of 732 · from 23 benchmark results across 5 sources Score within the [LMArena (text)](https://whatmodel.sh/?set=lmarena-text&score=set) comparison set only. ## What it's good at ● marks Step 3.5 Flash (196B-A11B) among every model on that task or test, worst to best. 
task            vs the others          rank  counts  tests
agent tasks     ▁▁▃▆█▄▅●▄▄▂▁     #84 of 313     13%      6
coding          ▁▁▁▃▅▇█▇●▄▃▁   #103 of 469*    3.3%      2
chat & writing  ▁▁▁▃▆▄▆█▆●▇▃   #125 of 402*    7.9%      1
math            ▁▁▃▅█▆▄●▄▂▁▁     #86 of 431     61%     13
web search      ██···█····█●       #1 of 5*    6.7%      3
agent tasks · #84 of 313 · 13% of its score Beats [GPT-5](gpt-5.md), [Grok 4.20 (0309 reasoning)](grok-4-20.md), [GPT-5.1](gpt-5-1.md) here, though they rank higher overall. Trails [MiniMax-M2.7](minimax-m2-7.md) here, though they rank lower overall. 
test                                result  vs the others in its run    difficulty           counts  why
BrowseComp-ZH (maker)                66.9%  ▁▁▁▁▁▃·▁▄▅▄● beats 19/23    hard · avg 20%         4.6%  at its level
τ²-bench aggregate (maker)           88.2%  ▄······▄█·▄● beats 5/5      medium · avg 51%       3.7%  typical
Terminal-Bench 2.0 (maker)           51.0%  ▂▂▅▅██●█▆▂▃▃ beats 20/36    very hard · avg 12%    3.1%  at its level · small run
BrowseComp (maker)                   51.6%  ▂·▃▁▂▂▆●█▂▂▄ beats 20/46    hard · avg 19%         2.1%  at its level · small run
SWE-bench Verified (maker)           74.4%  ▁▁▁▁▁▁▂▄█●▁▁ beats 102/138  hard · avg 19%            —  not counted yet: no other model in the …
BrowseComp with context managemen…   69.0%  ▄·▄●·█▂·▆▄·▂ beats 4/15                               —  not counted yet: no other model in the …
coding · #103 of 469 · 3.3% of its score Beats [Grok 4.20 (0309 reasoning)](grok-4-20.md), [Grok 4.3](grok-4-3.md), [Muse Glimmer 30B](muse-glimmer-30b.md) here, though they rank higher overall. Trails [Llama 3.1 Nemotron Ultra 253B v1](llama-3-1-nemotron-ultra-253b.md), [Qwen3 235B-A22B Thinking 2507](qwen3-235b-a22b-thinking-2507.md), [o4-mini](o4-mini.md) here, though they rank lower overall. 
test                                result  vs the others in its run    difficulty        counts  why
LiveCodeBench v6 (window unspecif…   86.4%  ▂▃▄▄▅██▆▄●▄▆ beats 56/67    medium · avg 48%    3.3%  small run
SWE-bench Verified (maker)           74.4%  ▁▁▁▁▁▁▂▄█●▁▁ beats 102/138  hard · avg 19%         —  not counted yet: no other model in the …
chat & writing · #125 of 402 · 7.9% of its score Beats [LongCat 2.0 (1.6T-A48B)](longcat-2-0.md) here, though they rank higher overall. Trails [LFM2.5 2.6B](lfm2-5-2-6b.md), [Nemotron-Cascade-8B](nemotron-cascade-8b.md), [DeepSeek V3 0324](deepseek-v3-0324.md) here, though they rank lower overall. 
test            result  vs the others in its run    difficulty  counts  why
LMArena (text)  1393.0  ▁▁▁▂▃▃▃▆▆●█▃ beats 237/385  avg 1260      7.9%  blunt
math · #86 of 431 · 61% of its score Beats [Claude Sonnet 4.6](claude-sonnet-4-6.md), [Claude Opus 4.5](claude-opus-4-5.md), [Qwen3.5 397B-A17B](qwen3-5-397b-a17b.md) here, though they rank higher overall. Trails [Gemini 2.5 Flash](gemini-2-5-flash.md), [Granite 4.2 30B](granite-4-2-30b.md), [Qwen3-Next 80B-A3B Instruct](qwen3-next-80b-a3b-instruct.md) here, though they rank lower overall. 
test                                result  vs the others in its run    difficulty            counts  why
MathArena ArXivMath April 2026 (4…   33.8%  ▁·····●██▄·▁ beats 1/20     very hard · avg 7.6%     16%  sharp · at its level
MathArena ArXivMath February 2026…   32.8%  ▁▁▁▁●▇▄▃▁▃·▃ beats 9/24     very hard · avg 9.4%    8.7%  at its level
MathArena ArXivMath March 2026 (3…   43.3%  ●▃▅·█▃▃·▅·▃▃ beats 0/14     hard · avg 16%          7.7%  at its level
MathArena ArXivMath January 2026 …   60.3%  ▂··▄▂▄█▆●▄▄█ beats 13/23    hard · avg 23%          6.8%  at its level
AIME 2025 (maker)                    97.3%  ▁·▁▁▁▂▅▆█▇▅● beats 133/133  easy · avg 69%          4.8%  sharp · maxed out · small run
MathArena ArXivMath December 2025…   41.9%  ▃·▃▃●▁·▄·▁▁▁ beats 10/18    hard · avg 19%          4.5%  at its level
MathArena BrokenArXiv April 2026 …   13.9%  ▂▂●█▃▃▂▂▂··▂ beats 2/18     very hard · avg 0.8%    3.1%  sharp · noisy
IMO-AnswerBench (maker)              85.4%  ▁▁·▁▁▁▄▆█●▂▃ beats 35/44    medium · avg 36%        2.6%  small run
MathArena BrokenArXiv March 2026 …    7.1%  ●·▆█▂·▂▂···▂ beats 3/14     very hard · avg 1.0%    1.8%  sharp · noisy
MathArena BrokenArXiv February 20…   11.3%  ▃▂▂▃●▃▂·▃··▂ beats 6/16     very hard · avg 0.9%    1.7%  sharp · noisy
MathArena ArXivLean March 2026 (4…    0.0%  ●····▂·····█ beats 0/5      very hard · avg 1.2%    1.7%  too hard for it
HMMT November 2025 (maker)           94.0%  ▁▁····▁▅▃█●▁ beats 19/23    easy · avg 66%          1.1%  noisy
HMMT February 2025 (maker)           98.4%  ▁▁▁▃▁▅▅▃██▁● beats 31/31    medium · avg 61%        0.8%  noisy · small run
web search · #1 of 5 · 6.7% of its score Beats [Kimi K2.5](kimi-k2-5.md) here, though they rank higher overall. 
test                                result  vs the others in its run  difficulty      counts  why
BrowseComp-ZH (maker)                66.9%  ▁▁▁▁▁▃·▁▄▅▄● beats 19/23  hard · avg 20%    4.6%  at its level
BrowseComp (maker)                   51.6%  ▂·▃▁▂▂▆●█▂▂▄ beats 20/46  hard · avg 19%    2.1%  at its level · small run
BrowseComp with context managemen…   69.0%  ▄·▄●·█▂·▆▄·▂ beats 4/15                        —  not counted yet: no other model in the …
other tests (7) 
test                                result  vs the others in its run  difficulty           counts  why
KINA (Knowledge Index of Noah's A…   29.1%  ▁▁▂▁▃▄●▆▄▅▂▃ beats 17/39  very hard · avg 21%    5.8%  at its level
WildClawBench                        26.7%  ▃·●▅█·▅█▅▃▃▃ beats 1/17   very hard · avg 11%    3.7%  at its level
ResearchRubrics (maker)              65.3%  ███··██·█··● beats 6/6    hard · avg 33%         3.1%  at its level
GAIA (no-file subset) (maker)        84.5%  ▃▃····█▃···● beats 6/6    medium · avg 49%       1.6%  at its level
BrowseComp-ZH with context manage…   73.7%  no other model yet                                  —  not counted yet: no other model in the …
xbench-DeepSearch (2025.05) (make…   83.7%  no other model yet                                  —  not counted yet: no other model in the …
xbench-DeepSearch (2025.10) (make…   56.3%  no other model yet                                  —  not counted yet: no other model in the …
## How the score was worked out Each result implies a score on its own, after adjusting for how hard that test is and who ran it. The model's 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 is added up by who ran the tests. [Method](../method.md) 
source                                    results  share of the score
MathArena ArXivMath / BrokenArXiv / ArX…        9   52% ▇▇▇▇▇▇
stepfun-ai/Step-3.5-Flash model card (s…       11   31% ▇▇▇▇
LMArena leaderboard dataset                     1    8% ▇
Knowledge Index of Noah's Ark                   1    6% ▇
WildClawBench: A Benchmark for Real-Wor…        1    4% ▇
prior: pull toward the middle                        0% ▇
weighted average                                   = 67.1
Licences: MathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasets ([CC BY-SA 4.0](../../licenses/matharena-sharealike-NOTICE.txt)) · stepfun-ai/Step-3.5-Flash model card ([Apache-2.0](../../licenses/stepfun-ai--Step-3.5-Flash.txt)) · LMArena leaderboard dataset ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · Knowledge Index of Noah's Ark ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · WildClawBench: A Benchmark for Real-World ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/))
Not counted (5) 
source                  test                            score  why
stepfun-ai/Step-3.5-F…  BrowseComp with context manag…  69.0%  too few models have it yet
stepfun-ai/Step-3.5-F…  BrowseComp-ZH with context ma…  73.7%  too few models have it yet
stepfun-ai/Step-3.5-F…  SWE-bench Verified              74.4%  too few models have it yet
stepfun-ai/Step-3.5-F…  xbench-DeepSearch (2025.05)     83.7%  too few models have it yet
stepfun-ai/Step-3.5-F…  xbench-DeepSearch (2025.10)     56.3%  too few models have it yet
## How to run it ### Through an API api (StepFun) · $0.15/M blended · $0.1 in / $0.3 out per 1M tokens · [platform.stepfun.ai](https://platform.stepfun.ai/docs/en/guides/models/step-3.5-flash.md) [price source](https://platform.stepfun.ai/docs/en/guides/pricing/details.md): USD per 1M tokens, StepFun Open Platform (international) step-3.5-flash: input on cache miss and output (output includes reasoning). Cache-hit input $0.02/1M excluded. Currency: USD. StepFun official pricing page, observed 2026-09-25. · [context source](https://github.com/anomalyco/models.dev/blob/a668908cfe7073faa0363699486fc41a3ff52412/models/stepfun/step-3.5-flash.toml) ### llama.cpp (recommended) 1. Install llama.cpp. Homebrew and winget installs name the commands llama-cli and llama-server instead. [Other ways to install](https://github.com/ggml-org/llama.cpp/blob/master/docs/install.md) ```sh # macOS / Linux curl -LsSf https://llama.app/install.sh | sh # Windows irm https://llama.app/install.ps1 | iex ``` 2. Run Step 3.5 Flash (196B-A11B) (Q4_K, needs ~118.1 GB, from [huggingface.co/ggml-org](https://huggingface.co/ggml-org/Step-3.5-Flash-GGUF/tree/main)) ```sh llama cli -hf ggml-org/Step-3.5-Flash-GGUF:Q4_K ``` 3. Or start a local server and open the address it prints (http://localhost:9931) ```sh llama serve -hf ggml-org/Step-3.5-Flash-GGUF:Q4_K ``` ### LM Studio 1. Install LM Studio. Open the app once before using the lms command. [Download it (macOS / Windows / Linux)](https://lmstudio.ai/download) 2. Run Step 3.5 Flash (196B-A11B) (Q4_K, needs ~118.1 GB, from [huggingface.co/ggml-org](https://huggingface.co/ggml-org/Step-3.5-Flash-GGUF/tree/main)) ```sh lms get https://huggingface.co/ggml-org/Step-3.5-Flash-GGUF@Q4_K -y ``` 3. Then run lms ls to see its name, and lms chat followed by that name Other versions: 4bit (~110.8 GB, Mac only) `lms get https://huggingface.co/mlx-community/Step-3.5-Flash-4bit -y` · 8bit (~209.3 GB, Mac only) `lms get https://huggingface.co/mlx-community/Step-3.5-Flash-8bit -y` ### MLX (Mac) 1. Install MLX (Mac). Apple silicon Macs only, text only. ```sh # with uv uv tool install mlx-lm # or pip install mlx-lm ``` 2. Run Step 3.5 Flash (196B-A11B) (4bit, needs ~110.8 GB, Mac only, from [huggingface.co/mlx-community](https://huggingface.co/mlx-community/Step-3.5-Flash-4bit/tree/caad23b0411c27d08aa3967822e61e5efcdd175c)) ```sh mlx_lm.chat --model mlx-community/Step-3.5-Flash-4bit ``` 3. Or start an OpenAI-compatible server at http://localhost:8080/v1 ```sh mlx_lm.server --model mlx-community/Step-3.5-Flash-4bit ``` Other versions: 8bit (~209.3 GB, Mac only) `mlx_lm.chat --model mlx-community/Step-3.5-Flash-8bit` Model repository and documentation: [huggingface.co/stepfun-ai](https://huggingface.co/stepfun-ai/Step-3.5-Flash) ## Published speed measurements Reference results for the hardware, engine and settings below. These are not verified speeds of the downloads above or predictions for your machine. 
hardware                      precision               speed  engine and settings
Apple Mac Studio M4 Max (12…  Q4_K_S (stepfun…  48.18 tok/s  StepFun llama.cpp fork (repositor…
Apple Mac Studio M4 Max (12…  Q4_K_S (stepfun…  45.96 tok/s  StepFun llama.cpp fork (repositor…
Apple Mac Studio M4 Max (12…  Q4_K_S (stepfun…   43.9 tok/s  StepFun llama.cpp fork (repositor…
Apple Mac Studio M4 Max (12…  Q4_K_S (stepfun…  41.27 tok/s  StepFun llama.cpp fork (repositor…
Apple Mac Studio M4 Max (12…  Q4_K_S (stepfun…  38.22 tok/s  StepFun llama.cpp fork (repositor…
Apple Mac Studio M4 Max (12…  Q4_K_S (stepfun…  31.52 tok/s  StepFun llama.cpp fork (repositor…
Apple Mac Studio M4 Max (12…  Q4_K_S (stepfun…   8.88 tok/s  StepFun llama.cpp fork (repositor…
NVIDIA DGX Spark (GB10)       Q4_K_S (stepfun…  19.79 tok/s  StepFun llama.cpp fork (repositor…
NVIDIA DGX Spark (GB10)       Q4_K_S (stepfun…  20.27 tok/s  StepFun llama.cpp fork (repositor…
NVIDIA DGX Spark (GB10)       Q4_K_S (stepfun…  20.25 tok/s  StepFun llama.cpp fork (repositor…
NVIDIA DGX Spark (GB10)       Q4_K_S (stepfun…  19.24 tok/s  StepFun llama.cpp fork (repositor…
NVIDIA DGX Spark (GB10)       Q4_K_S (stepfun…  17.75 tok/s  StepFun llama.cpp fork (repositor…
NVIDIA DGX Spark (GB10)       Q4_K_S (stepfun…  15.42 tok/s  StepFun llama.cpp fork (repositor…
NVIDIA DGX Spark (GB10)       Q4_K_S (stepfun…   8.39 tok/s  StepFun llama.cpp fork (repositor…
AMD Ryzen AI Max+ 395 (Stri…  Q4_K_S (stepfun…   3.43 tok/s  StepFun llama.cpp fork (repositor…
AMD Ryzen AI Max+ 395 (Stri…  Q4_K_S (stepfun…    3.4 tok/s  StepFun llama.cpp fork (repositor…
AMD Ryzen AI Max+ 395 (Stri…  Q4_K_S (stepfun…   3.36 tok/s  StepFun llama.cpp fork (repositor…
AMD Ryzen AI Max+ 395 (Stri…  Q4_K_S (stepfun…    3.3 tok/s  StepFun llama.cpp fork (repositor…
AMD Ryzen AI Max+ 395 (Stri…  Q4_K_S (stepfun…   3.19 tok/s  StepFun llama.cpp fork (repositor…
AMD Ryzen AI Max+ 395 (Stri…  Q4_K_S (stepfun…   2.98 tok/s  StepFun llama.cpp fork (repositor…
Each speed links to its source, which has the full settings. [interactive version](https://whatmodel.sh/model/step-3-5-flash) · [JSON](https://whatmodel.sh/api?format=json&all=1)