Step 3.7 Flash

2026-05-29 · Publisher : StepFun · Domain : Artificial Intelligence

StepFun releases Step 3.7 Flash, a vision-language mixture-of-experts model with 198 billion parameters of which 11 billion are active, under Apache 2.0, with 256,000 tokens of context and up to 400 tokens per second.

### Context Step 3.7 Flash arrives four months after Step 3.5 Flash. It adds native vision to the same 196-billion-parameter language base and targets agents that combine perception, search and reasoning, with frequent production workloads: reading financial reports, search loops and concurrent coding agents. ### What the model brings - Native vision: 1.8-billion-parameter encoder; 79.2 on SimpleVQA (Search), first rank according to the model card, and 95.3 on V* (Python), state-of-the-art level. - Three levels of extended thinking (low, medium, high) to trade off speed, cost and depth. - Agents: 67.1 on ClawEval-1.1, ahead of the next competitor (59.8 according to the model card), 49.5 on Toolathlon and 48.1 on Humanity's Last Exam with tools. - Coding: 56.3 on SWE-Bench Pro, second rank in the model card's selection; Terminal-Bench 2.1 at 59.5 and GDPVal-AA at 45.8 are flagged as areas for improvement. - “Advisor” mode described in the post: a lightweight executor consults a frontier model at critical moments, at a cost of $0.19 per task versus $1.76 for Claude Opus 4.6 on SWE-Bench Verified (vendor claim). ### Architecture and specifications - Architecture: vision-language MoE with 198 billion parameters (196 billion language and 1.8 billion vision), about 11 billion active. - Context: 256,000 tokens. Throughput: up to 400 tokens per second. Modalities: text and image. - License: Apache 2.0 (free commercial use). - Weights: stepfun-ai/Step-3.7-Flash on Hugging Face (and GGUF); API on platform.stepfun.ai, OpenRouter and NVIDIA NIM. Knowledge cutoff: not disclosed. ### Cost and access Pricing of StepFun's official API, according to the model card: $0.20 per million input tokens (cache miss), $0.04 for input with cache and $1.15 for output. For self-hosting, StepFun cites unified-memory machines with at least 128 GB (Mac Studio, MacBook Pro, AMD Ryzen AI Max+ 395, NVIDIA DGX Station) and data-center deployment. | Item | Price (USD per million tokens) | | --- | --- | | Input (cache miss) | 0.20 | | Input (cache hit) | 0.04 | | Output | 1.15 | ### Benchmarks | Metric | Step 3.7 Flash | Published rank or comparison | | --- | --- | --- | | SWE-Bench Pro | 56.3 | Second rank in the selection; GPT-5.5: 58.6 (post, via page summary) | | ClawEval-1.1 | 67.1 | First rank; next competitor: 59.8 | | SimpleVQA (Search) | 79.2 | First rank | | V* (Python) | 95.3 | Parity with the frontier | | Terminal-Bench 2.1 | 59.5 | Flagged area for improvement | | GDPVal-AA | 45.8 | Flagged area for improvement | | Toolathlon | 49.5 | Not specified | | Humanity's Last Exam (with tools) | 48.1 | Not specified | Source: Step 3.7 Flash Hugging Face model card (vendor scores); comparison with GPT-5.5 taken from the StepFun post based on a page summary. No closed-model table is reproduced in full for lack of a stable source. ### Versus closed models The only comparison with a closed model retained here is SWE-Bench Pro: 56.3 for Step 3.7 Flash versus 58.6 for GPT-5.5, a gap of 2.3 points, with the GPT-5.5 value cross-checked against the LongCat-2.0 model card. On Terminal-Bench 2.1, Step 3.7 Flash at 59.5 is clearly behind GPT-5.5 (73.8 in the LongCat-2.0 model card, a value cited from an official report), and StepFun acknowledges it as an area for improvement. The model card's top scores (ClawEval-1.1, SimpleVQA) are based on a cohort chosen by the vendor, not an independent ranking. The output price of $1.15 per million tokens should be compared with the pricing of frontier closed models, which are not recorded in this page. For an open-or-closed choice, the strength is a lightweight multimodal model (11 billion active parameters) under Apache 2.0, deployable on machines with 128 GB of memory. ### Key takeaways Step 3.7 Flash adds vision and extended-thinking levels to a lightweight MoE, with an output price of $1.15. It is close to closed models on SWE-Bench Pro but clearly behind on Terminal-Bench, according to figures from the vendor and a cited third party.

Tags : LLM, Open Source, StepFun, Step, Apache 2.0, MoE, Multimodal