--- title: "StepFun's Step 3.7 Flash Hits 416 Tokens per Second While Staying Open-Source" subtitle: "StepFun's open-weights 198B MoE hits 400+ tokens per second with only 11B active parameters, multi-token prediction, and native vision input." url: "https://alphasignal.ai/news/stepfun-s-step-3-7-flash-hits-416-tokens-per-second-while-staying-open-source" author: "AlphaSignal Newsroom" author_url: "https://alphasignal.ai/editorial-team" company: "Artificial Analysis" company_url: "https://alphasignal.ai/company/artificialanlys" published: "2026-06-04T03:48:53.000Z" updated: "2026-06-04T23:08:38.860Z" type: "News" topics: - "Llms" - "Open Source" subtopics: - "Mixture Of Experts" - "Small Models" upvotes: 238 read_time: "6 min" source_url: "https://artificialanalysis.ai/models/step-3-7-flash" comments: 0 --- # StepFun's Step 3.7 Flash Hits 416 Tokens per Second While Staying Open-Source > StepFun's open-weights 198B MoE hits 400+ tokens per second with only 11B active parameters, multi-token prediction, and native vision input. StepFun has pushed a new open-weights model into the Pareto frontier of speed versus intelligence. [Step 3.7 Flash](https://huggingface.co/stepfun-ai/Step-3.7-Flash) is a 198B-parameter sparse Mixture-of-Experts vision-language model that activates roughly 11B parameters per token, scores 43 on the Artificial Analysis Intelligence Index, and serves at over 400 output tokens per second on StepFun's first-party API. It ships under Apache 2.0, with weights in BF16, FP8, NVFP4, and GGUF formats, putting frontier-adjacent capability inside the memory budget of a single high-end workstation. ## The speed number is the headline According to [Artificial Analysis](https://artificialanalysis.ai/models/step-3-7-flash), Step 3.7 Flash generates output at 415.9 tokens per second based on StepFun's API, which is well above average compared to other open weight models of similar size (median: 57.4 t/s). Time to first token sits at 0.96 seconds, against a median of 2.36 seconds for comparable models. That places it at #1 of 88 models tracked for output speed. Two architectural choices drive that throughput. First, only about 11B of the 198B parameters fire per token thanks to MoE routing. Second, the model ships with trained Multi-Token Prediction heads, three of them, that predict several future tokens in a single forward pass. Combined with speculative decoding, the model can verify multiple draft tokens at once instead of generating them one at a time. The reference SGLang serving config exposes this directly with flags like `--speculative-algorithm EAGLE`, `--speculative-num-steps 3`, and `--enable-multi-layer-eagle`. ## What the intelligence numbers actually say On the Artificial Analysis Intelligence Index v4.0, which blends ten evaluations including GDPval-AA, Terminal-Bench Hard, SciCode, AA-LCR, Humanity's Last Exam, and GPQA Diamond, Step 3.7 Flash lands at 43. That's a four-point gain over Step 3.5 Flash 2603 (38.5), roughly matching Qwen3.5 122B A10B (41.6) but trailing MiniMax-M2.7 (49.6) and DeepSeek V4 Flash (46.5). The biggest movements come from agentic tasks: - GDPval-AA Elo rises from 1070 to 1298 - Terminal-Bench Hard goes from 32.6% to 35.6% - AA-LCR long-context reasoning jumps from 54.3% to 63.7% - MMMU-Pro multimodal reasoning lands at 75.3%, essentially tied with Qwen3.5 122B A10B at 75.0% StepFun's own benchmarks push the agentic story harder. On SWE-Bench Pro, Step 3.7 Flash scores 56.26%, up from Step 3.5 Flash's 51.3%. On Terminal-Bench 2.1, it scores 59.55%, up from 53.37%. Cross-harness consistency on StepFun's internal Step-SWE-Bench shows Step 3.5 Flash ranging from 43% to 73% across harnesses, while Step 3.7 Flash ranges from 64.5% to 71.5%. Narrower variance across scaffolds matters a lot when you're plugging the same model into half a dozen different agent frameworks. ## Where it falls short Knowledge breadth is the soft spot. Step 3.7 Flash posts an AA-Omniscience score of -38 with a 25.4% accuracy and an 84.4% hallucination rate, still trailing peers like MiniMax-M2.7 and DeepSeek V4 Flash, although it did improve from Step 3.5 Flash's -44. It's also verbose: Artificial Analysis notes that when evaluated on the Intelligence Index, Step 3.7 Flash generated 260M output tokens, which is at the higher end compared to other open weight models of similar size (median: 43M). Fast tokens are still tokens you pay for. ## The Advisor Mode trick StepFun is leaning into a hybrid pattern borrowed from Anthropic's playbook. Step 3.7 Flash supports Advisor Mode, where the model runs the agentic loop end-to-end, calling tools, reading results, iterating, and escalates to a larger advisor model only at specific inflection points such as planning or recovering from repeated failures. The pitch is that you only pay frontier prices at the moments that actually need frontier reasoning. StepFun's vendor-reported numbers for this setup are aggressive. With Advisor Mode enabled on SWE-Bench Verified, StepFun reports Step 3.7 Flash reaches 97% of Claude Opus 4.6's coding performance at roughly one-ninth the per-task cost ($0.19 vs. $1.76 per task). These are StepFun's internal figures. Treat them with the usual caution, but the underlying arithmetic, $0.20 per million input tokens and $1.15 per million output tokens with an 80% cache-hit discount, is real. ## Vision lands in the Flash line Where Step 3.5 Flash was text-only, Step 3.7 Flash introduces a 1.8B-parameter Vision Transformer for native image understanding. For recognition tasks where the model's parametric knowledge is insufficient, such as long-tail entities or recently emerged concepts, it invokes a visual search tool to retrieve and verify; on SimpleVQA with Search it scores 79.16%, comparable to GPT 5.5 at 79.11%. For fine-grained visual tasks it uses a Python code interface to crop, zoom, and draw bounding boxes; on V\* with Python it scores 95.29%, and on HR-Bench 4K and 8K it scores 89.13% and 86.34% respectively. Among same-size open-weights peers, MiniMax-M2.7, DeepSeek V4 Flash, and gpt-oss-120b remain text-only. ## Hardware and deployment At BF16 native precision, the weights occupy ~400GB. StepFun also ships FP8 (~200GB) and NVFP4 (~100GB) variants, plus GGUF builds for llama.cpp. The Hugging Face card spells out the local-deployment floor: GGUF runtime overhead of about 7 GB, minimum unified memory or VRAM of 120 GB (such as Mac Studio, NVIDIA DGX Station, AMD Ryzen AI Max+ 395), with 128 GB recommended. A hands-on review running the IQ4\_XS quantization (105GB) on an NVIDIA DGX Spark with 128GB unified memory using llama.cpp on StepFun's custom step3.7 branch reports total time from download to first inference under 30 minutes. Distribution is broad from day one. Step 3.7 Flash is available on the StepFun Open Platform (platform.stepfun.ai globally, platform.stepfun.com in China), OpenRouter, and NVIDIA NIM, with DeepInfra, Fireworks AI, and Modal coming soon. Quick API usage looks like: ``` from openai import OpenAI import os client = OpenAI( api_key=os.environ["STEP_API_KEY"], base_url="https://api.stepfun.ai/v1", ) completion = client.chat.completions.create( model="step-3.7-flash", messages=[{"role": "user", "content": "Summarize this repo"}], ) ``` ## Why this release matters The competitive frame for open weights has shifted from raw benchmark scores to cost per agent loop. StepFun is one of China's "AI Six Tigers," and the company has reportedly closed a pre-IPO round of roughly $2.5B at a $10B post-money valuation, which helps explain the cadence and the open-weight strategy. The bet is that teams running thousands of agent runs per month care less about the top of GPQA Diamond and more about whether they can keep an agent loop at $0.20 per task without losing reliability across scaffolds. With three reasoning levels, native multimodal input, MTP-accelerated decoding, and a permissive license that lets you self-host on a single Mac Studio or DGX Spark, Step 3.7 Flash is one of the cleaner answers to that question shipped this year. For tool-heavy coding agents, browser agents, or vision-grounded UI automation, it's an obvious model to benchmark against your current default. ## Sources - [X post by @artificialanlys](https://x.com/artificialanlys/status/2062381047212638697) - [Step 3.7 Flash - Intelligence, Performance & Price Analysis](https://artificialanalysis.ai/models/step-3-7-flash) - [stepfun-ai/Step-3.7-Flash - Hugging Face](https://huggingface.co/stepfun-ai/Step-3.7-Flash) - [StepFun Releases Step 3.7 Flash: A 198B MoE Vision-Language Model](https://www.marktechpost.com/2026/05/29/stepfun-releases-step-3-7-flash-a-198b-moe-vision-language-model-for-coding-agents-and-search-workflows/) - [StepFun Step 3.7 Flash: 196B MoE Agentic Vision Model](https://www.digitalapplied.com/blog/stepfun-step-3-7-flash-196b-moe-agentic-vision-release) - [Step 3.7 Flash Review: Tested on a DGX Spark](https://flowtivity.ai/blog/step-3-7-flash-review-dgx-spark/)