StepFun releases Step 3.7 Flash, an open-weight 198B MoE model that runs at 400 tokens per second

VIEWS44.6KBOOKMARKS87LIKES450RETWEETS22REPLIES28

Our users love @StepFun_ai models and this new release packs a punch at a small size. Looking forward to seeing how well it works with Hermes Agent!

StepFun@StepFun_ai

⚡️ Step 3.7 Flash is here: The new frontier is agent efficiency.

#1 ClawEval-1.1 (67.1), #1 SimpleVQA Search (79.2), #2 SWE-PRO (56.3), 95.3 on V* Python. Open weights under Apache 2.0.

Built for agentic, coding, search, and multimodal workflows — balancing speed, cost, and reliable execution.

- 400 TPS. 198B sparse MoE, ~11B active. 256K context, 3 reasoning levels. - Understands UIs, charts, docs, images — then writes code or calls tools to act on what it sees. - Web + visual search reaches further: more sources, deeper follow-up. - Reliable tool use — less drift, fewer broken toolcalls. 98%+ on τ²-bench across all difficulty levels. - Works with Claude Code, KiloCode, Hermes Agent, OpenClaw, and protocols like MCP. - Runs locally on Mac Studio M4 Max, DGX Spark, AMD AI Max+ 395.

GitHub: http://github.com/stepfun-ai/Step-3.7-Flash HuggingFace: http://huggingface.co/stepfun-ai/Step-3.7-Flash GGUF: http://huggingface.co/stepfun-ai/Step-3.7-Flash-GGUF ModelScope: http://modelscope.cn/models/stepfun-ai/Step-3.7-Flash API: http://platform.stepfun.ai Blog: http://static.stepfun.com/blog/step-3.7-flash/

26d44.6K45087

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

I've been waiting for this! They managed to do it before June, and they open sourced it right away! @antirez I've been saying. Look at this model. It's much smaller than V4-Flash, it's multimodal, it's fast. It deserves to be added.

StepFun@StepFun_ai

⚡️ Step 3.7 Flash is here: The new frontier is agent efficiency.

#1 ClawEval-1.1 (67.1), #1 SimpleVQA Search (79.2), #2 SWE-PRO (56.3), 95.3 on V* Python. Open weights under Apache 2.0.

Built for agentic, coding, search, and multimodal workflows — balancing speed, cost, and reliable execution.

- 400 TPS. 198B sparse MoE, ~11B active. 256K context, 3 reasoning levels. - Understands UIs, charts, docs, images — then writes code or calls tools to act on what it sees. - Web + visual search reaches further: more sources, deeper follow-up. - Reliable tool use — less drift, fewer broken toolcalls. 98%+ on τ²-bench across all difficulty levels. - Works with Claude Code, KiloCode, Hermes Agent, OpenClaw, and protocols like MCP. - Runs locally on Mac Studio M4 Max, DGX Spark, AMD AI Max+ 395.

GitHub: http://github.com/stepfun-ai/Step-3.7-Flash HuggingFace: http://huggingface.co/stepfun-ai/Step-3.7-Flash GGUF: http://huggingface.co/stepfun-ai/Step-3.7-Flash-GGUF ModelScope: http://modelscope.cn/models/stepfun-ai/Step-3.7-Flash API: http://platform.stepfun.ai Blog: http://static.stepfun.com/blog/step-3.7-flash/

26d15.2K25359

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

Step 3.7 is generally neck and neck with V4-Flash (which is underappreciated as a powerful agent), I think they targeted it. But it goes to show that vision is a must now. V4.1V can't come soon enough.

26d13.4K25843

Niels Rogge@NielsRogge

Impressive release by StepFun, explore it at https://paperswithcode.co/paper/83892

StepFun@StepFun_ai

⚡️ Step 3.7 Flash is here: The new frontier is agent efficiency.

#1 ClawEval-1.1 (67.1), #1 SimpleVQA Search (79.2), #2 SWE-PRO (56.3), 95.3 on V* Python. Open weights under Apache 2.0.

Built for agentic, coding, search, and multimodal workflows — balancing speed, cost, and reliable execution.

- 400 TPS. 198B sparse MoE, ~11B active. 256K context, 3 reasoning levels. - Understands UIs, charts, docs, images — then writes code or calls tools to act on what it sees. - Web + visual search reaches further: more sources, deeper follow-up. - Reliable tool use — less drift, fewer broken toolcalls. 98%+ on τ²-bench across all difficulty levels. - Works with Claude Code, KiloCode, Hermes Agent, OpenClaw, and protocols like MCP. - Runs locally on Mac Studio M4 Max, DGX Spark, AMD AI Max+ 395.

GitHub: http://github.com/stepfun-ai/Step-3.7-Flash HuggingFace: http://huggingface.co/stepfun-ai/Step-3.7-Flash GGUF: http://huggingface.co/stepfun-ai/Step-3.7-Flash-GGUF ModelScope: http://modelscope.cn/models/stepfun-ai/Step-3.7-Flash API: http://platform.stepfun.ai Blog: http://static.stepfun.com/blog/step-3.7-flash/

26d18.4K4513

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

First impressions: StepFun 3.7 vision is kinda low-res and hallucinatory, behind MiMo 2.5. Kimi >> DS-Vision > MiMo > StepFun. Well, it's their first vision in the Flash series and this is by far the smallest and fastest model. But no bueno.

26d3.9K655

Niels Rogge@NielsRogge

Small new feature btw, you can now hover over all models on a given leaderboard! 🔥

Let me know which features you'd like to see next!

Try it out here, for example: https://paperswithcode.co/benchmark/mmmu

Niels Rogge@NielsRogge

Impressive release by StepFun, explore it at https://paperswithcode.co/paper/83892

26d8.9K229

嫣彤~🌸同城上门@Beatric93216624

@NtRqC21USkYxjNs @StepFun_ai 👆‌🔥

💖 🙌 ⭐😋

26d3

雨柔~🌸同城上门@Billy0683925968

@NtRqC21USkYxjNs @StepFun_ai 👆✌

🙏‍

👍 ‍👏😉

26d3

洛妤~🌸同城上门@Charles12315406

@NtRqC21USkYxjNs @StepFun_ai 👆✌

⭐

🤝 👏

26d2

0xSero@0xSero

@StepFun_ai @scaling01 FP8 and NVFP4 cooking rn

26d1.4K261

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

and I lowkey respect them not going for DeepSeek prices like MiMo, lol. They have the competitive edge in multimodality + decode speed, but more importantly they *can't afford* to reduce cache hit costs by 11x. Architecturally, it's almost identical to MiMo.

26d1.7K51

Anurag Bhagsain@abhagsain

@StepFun_ai Step 3.7 Flash could be a great sub-agent model Much faster and almost as good as Gemini Flash 3.0 at half the cost Struggling a bit with the tool calling on my limited evals, but I think it should work with better prompting

26d82862

Artillain ( Mr. Zerolith )@artillain

@T_Wrex_Baby @StepFun_ai Disagree, most LLMs get dumber after 100k, if you're running agents that break up context, you are getting smarter output with a smalller context window vs a big one

This is one of the reasons ralph wiggum is so popular, even with claude users

26d142101

0xSero@0xSero

@teortaxesTex I agree the only reason ds4 isn’t running 24/7 on machine is because I need vision models sometimes.

I have loved step-3.5-flash

26d96342

Nous Research@NousResearch

@StepFun_ai Congrats on the release, the model looks great!

26d749141

Alaa Taieb@alaa_taieb

@StepFun_ai What models are you comparing this to? The graph is lacking labeling

26d87612

Louis@Louis9687221579

@teortaxesTex they cook, deepfried, burn deepseek V4 flash with just 70% of the size + RL over 3.5 baseline. Also seem like everyone got memory hole about "3.6" release. And they're also exploring larger models as well comments from https://www.reddit.com/r/LocalLLaMA/comments/1r8snay/ama_with_stepfun_ai_ask_us_anything/?show=original

26d5271

Sudo su@sudoingX

@NousResearch @Teknium @StepFun_ai will report back soon

26d38441

Conor@jconorgrogan

@StepFun_ai this model is very very very good for its size! well done.

26d8202

Nilay@localm_tuts

@StepFun_ai quick question - how can it work with single dgx when you have 110gb + weights.

26d7602