Step 3.5 Flash

2026-01-29 · Publisher : StepFun · Domain : Artificial Intelligence

StepFun has opened Step 3.5 Flash, a mixture-of-experts model of 196 billion parameters with 11 billion active, under Apache 2.0, with 256,000 tokens of context and generation of 100 to 300 tokens per second.

### Context Step 3.5 Flash succeeds Step3 (July 2025). In early 2026, the open references are DeepSeek V3.2, Kimi K2.5 and GLM-4.7, and the closed references are Claude Opus 4.5, GPT-5.2 and Gemini 3 Pro, which appear in StepFun's table. The model targets the fast agent: deep reasoning with low latency. ### What the model brings - Speed: prediction of 4 tokens per pass (MTP-3) for a throughput of 100 to 300 tokens per second in typical use and up to 350 in single-stream coding. - Agents and code: 74.4% on SWE-bench Verified and 51.0% on Terminal-Bench 2.0, with a scalable reinforcement learning framework. - Long context at reduced cost: sliding-window attention with 3 sliding layers for every 1 full-attention layer, for 256,000 tokens. - Local execution: StepFun cites the Mac Studio M4 Max and the NVIDIA DGX Spark; in INT4 on a 128 GB DGX Spark, the model generates about 20 tokens per second with 256,000 tokens of context. - Parallel thinking mode: the model card shows higher scores with parallel thinking (faded bars in the chart). ### Architecture and specifications - Architecture: MoE of 196.81 billion parameters in total, about 11 billion active; 45 layers, hidden dimension of 4,096, 288 routed experts and 1 shared, 8 experts per token. - Context: 256,000 tokens. Vocabulary: 128,896 tokens. Modalities: text. MTP head: sliding-window attention and dense FFN. - License: Apache 2.0 (free commercial use). - Weights: stepfun-ai/Step-3.5-Flash on Hugging Face and ModelScope; API on platform.stepfun.ai and OpenRouter. Knowledge cutoff: not disclosed. ### Cost and access According to aggregators, the API is priced at about $0.10 per million input tokens and $0.30 per million output tokens ($0.09 input at the cheapest of the providers); the official pricing page was not consulted. For self-hosting, StepFun cites a Mac Studio M4 Max and a 128 GB NVIDIA DGX Spark (INT4, about 20 tokens per second). | Item | Price / configuration | | --- | --- | | Input (aggregators, USD per million tokens) | 0.10 (from 0.09 depending on the provider) | | Output (aggregators, USD per million tokens) | 0.30 | | Local | 128 GB DGX Spark in INT4: about 20 tokens/s at 256,000 context | ### Benchmarks | Benchmark | Step 3.5 Flash | Claude Opus 4.5 | GPT-5.2 xhigh | Gemini 3.0 Pro | | --- | --- | --- | --- | --- | | Average of the blog post's selection | 81.0 | 80.6 | 82.2 | 80.7 | | AIME 2025 | 97.3 | 92.8 | 100.0 | 95.0 | | HMMT 2025 (average) | 96.2 | 92.3 | 98.3 | 96.0 | | IMOAnswerBench | 85.4 | 84.0 | 86.3 | 83.3 | | LiveCodeBench-V6 | 86.4 | 84.8 | 87.7 | 90.7 | | SWE-bench Verified | 74.4 | 80.9 | 80.0 | 76.2 | | τ²-Bench | 88.2 | 92.5 | 85.5 | 90.7 | Source: official StepFun blog post (comparison table with closed models), vendor scores. The comparisons with DeepSeek V3.2, Kimi K2.5, GLM-4.7, MiniMax M2.1 and MiMo-V2 Flash (open) are in the Hugging Face model card. ### Versus closed models On mathematical reasoning, Step 3.5 Flash sits between Claude Opus 4.5 and GPT-5.2 xhigh (AIME 2025: 97.3 versus 92.8 and 100.0) and ahead of Gemini 3.0 Pro on AIME and IMOAnswerBench. It is behind on software engineering (SWE-bench Verified 74.4 versus 80.9 for Opus 4.5 and 80.0 for GPT-5.2), on competitive coding against Gemini 3.0 Pro (86.4 versus 90.7) and on τ²-Bench against Opus 4.5 (88.2 versus 92.5). The average of the blog post's selection is 81.0, against 80.6 to 82.2 for the three closed models: it is an average chosen by the vendor, not an independent ranking. The price gap is large if the aggregators' pricing is accurate ($0.10 and $0.30 per million tokens), but the pricing of the three closed models is not recorded here. The model can be used locally on high-end hardware, which matters for use cases where data must not leave the company. ### Key takeaways Step 3.5 Flash shows that an MoE with 11 billion active parameters can approach the closed models of early 2026 in reasoning, with an Apache 2.0 license and possible local deployment. The gap persists on software engineering, according to the vendor's figures.

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