StepFun has released the weights of its flagship Step-5-Preview model: the TypeSafeAI/Step-5-Preview-BF16 checkpoint appeared on Hugging Face on September 20, 2026. It is a sparse Mixture-of-Experts with 600B total parameters, of which only 27B are active per token, i.e., about 4.5%. Context reaches 1M tokens, accepting text, images, and video as input. The company claims a place in the top three open-weight models and positions the release as a step toward efficiency in converting compute into intelligence. If independent checks confirm the numbers, this will change the economics of both the open segment and closed APIs.
What happened
On September 20, 2026, the open checkpoint TypeSafeAI/Step-5-Preview-BF16 with the weights of StepFun's flagship model appeared on Hugging Face. The architecture is a sparse Mixture-of-Experts on 92 layers of a "narrow and deep" scheme with Sparse GQA and block-wise token merging: 600B total parameters, of which 27B are active per token, approximately 4.5%. The cost of indexing long context in the card is stated as approximately one-eighth of a dense baseline, and the context itself reaches 1M tokens. It accepts text, images, and video in MP4, QuickTime, and Matroska formats up to 128 MB in size. The card claims an Intelligence Index of 44 according to Artificial Analysis v4.3.2 — top 3 among open-weight models, as well as 67.7 on DeepSWE v1.1 and 33.3 on Terminal-Bench v4, which is 2.6 times higher than Kimi K3's result. Generation speed is 99.8 tokens per second with a TTFT of 2.96 seconds. It also includes demonstrative demos: autonomous 22-hour optimization of a CUDA kernel for H100 to 508 TFLOPS versus 493 for Claude Opus 5, and independent post-training of Qwen3-30B-A3B, which raised the AIME24 result from 53.3% to 60% in 24 hours.
Context
Sparse Mixture-of-Experts have already become the standard for large open models, but usually a high total capacity comes with a relatively large active part. The closest benchmark in the open-weight segment is Kimi K3 Max, carrying about 2.8T parameters; StepFun, in response, bets not on size, but on the minimal share of active parameters per token. The release numbering also speaks to the intentions: the company skipped the entire Step 4.x line and jumped from Step-3.7-Flash directly to Step 5, promoting the "Pareto Frontier" philosophy — efficiency in converting compute into intelligence instead of simply scaling up. In this logic, opening the weights is not just a gesture of openness, but a way to verify: any numbers from the card can be reproduced by the community on their own hardware, which is not possible with closed models.
Why this matters for the industry
For the industry, the main thing here is architectural economics, not benchmarks. With 27B active parameters per token, near-frontier quality is inferred on 8× H100 in BF16, and an FP8 variant on 4× H100 is also claimed. This is direct price pressure on closed models like Claude Opus 5 and GPT-6 Astra: if the open stack provides comparable quality cheaper, the pricing power of API vendors relies more and more on inertia than on access price. Within the open-weight segment, the coordinate system is changing: they compare not the total size, but the cost of active capacity per token, and the "600B/27B" configuration, if the claimed results are confirmed, will become an efficiency benchmark. Startups have an open near-frontier candidate with multimodal input that can be deployed in their own environment and run through their own eval network; if the "Pareto Frontier" philosophy takes hold, competition will shift from a race of total parameters to a race of active capacity per token.
Why this matters for users
The weights are available for download, so any claim from the card can be verified on your own tasks. Local launch is supported by vLLM and SGLang — for example, with the vllm serve command with the --reasoning-parser stepfun flag. Your own hardware is not required: the model is called via an OpenAI-compatible API at api.stepfun.com under the name step-5-preview. For integrations, parallel tool calls, strict JSON Schema, and reasoning_effort levels low, medium, high, and xhigh are important — the familiar agent contract of closed APIs is transferred without rewriting pipelines. A million tokens of context and video input allow trying scenarios that previously required expensive narrow solutions: working with very long documents and analyzing video files. A reasonable first step is to compare step-5-preview with your current closed vendor on real tasks and measure not the claimed numbers, but the price per solved task.
What is still unknown / limitations
All key quality claims — Intelligence Index 44, DeepSWE v1.1 67.7, Terminal-Bench v4 33.3 — are taken from the model card and have not yet been confirmed by independent runs; the benchmark sample is narrow, there is no unified methodology for comparison with closed-source models, which leaves the risk of cherry-picking. The demonstrative demos with the CUDA kernel and Qwen3-30B-A3B post-training are provided without methodologies, seeds, and environment descriptions, so they are correctly considered a marketing demonstration, not proof. Hosted-API pricing is not disclosed in the available materials. A separate caveat: the author's assessment of the Telegram channel that up to 70% of the profit in NVFP4 quantization comes from MoE sparsity — this is a personal assessment, not StepFun data; it mixes the contribution of quantization and the contribution of sparsity, and without a controlled measurement of BF16 versus NVFP4 on the same hardware and traffic, it is not verifiable.
Sources
Author
Look at AI, editorial team