StepFun Launches Step 5 Preview: A 600B-Total, 27B-Active MoE Model With 1M Context for Long-Horizon Agentic Work
MarkTechPost Michal Sutter
StepFun launched Step 5 Preview, a big MoE model for coding, knowledge work, and finance. It’s built around a 1M-token context and cheaper API pricing, but open weights are still months away.
Based on reporting by MarkTechPost, Michal Sutter — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
StepFun has put out Step 5 Preview, its new flagship model for agentic work. The company is aiming it at software engineering, professional knowledge work, and finance, and the pitch is blunt: similar intelligence, lower task cost. That’s the whole “Pareto frontier” idea behind the launch name. The model is already usable through a hosted API and on StepFun’s platform. Self-hosting is still off the table until open weights arrive on October 15, 2026.
The model is a sparse mixture-of-experts system with about 600B total parameters and roughly 27B active per token. StepFun’s own docs list a 1M-token context window, text, image, and video input, text output, and support for low, medium, and high reasoning effort. It also supports streaming, tool calling, JSON Mode, JSON Schema, and prompt caching. On research tasks, the company says it handled 950 web fetches in a single agent action, and it also points to a Claude Code integration through Step Plan.
Architecturally, StepFun went narrow and deep rather than wide. Pandaily says the model uses 92 Transformer layers, and the research team argues that deeper stacks help with multi-hop reasoning over long stretches of prefill, especially when an agent is searching, running code, and reading tool output. Training also leans on on-policy, long-horizon reinforcement learning. StepFun says it uses bit-wise train and inference alignment across MoE routing, plus MTP-3 speculative decoding, FP8 MoE, and KV-cache offload. The company claims more than 3x end-to-end speedup for long-horizon RL.
On benchmarks, the company’s own numbers are mixed. StepFun says Step 5 Preview hit 66.4 on FrontierFinance and 83.3 on DRACO, while Claude Opus 5 and GPT-6 Astra did better on those tests. On coding, StepFun reports 67.7 on DeepSWE v1.1, 49.0 on StepCodeBench, and 80.5 on ProgramBench, with GPT-6 Astra and Claude Opus 5 ahead on all three. The company also describes two 24-hour agent runs: one where the model tuned an H100 kernel to 508 TFLOPS, and another where it pushed Qwen3-30B-A3B on AIME24 from 53.3% to 60% through automated post-training.
Independent numbers are a bit less dramatic, but still strong for the price. Artificial Analysis scores Step 5 Preview at 44 on its Intelligence Index, above the 24 median it cites for similar-priced reasoning models. It also measured output at 99.8 tokens per second on StepFun’s API. The API lists pricing at $1.00 per 1M input tokens on cache miss, $0.05 on cache hit, and $2.70 per 1M output tokens including reasoning. That looks cheap on paper, though the model’s 160M output tokens in the index run suggest verbose reasoning can chew through the savings fast.
My take — AI-written commentary, not fact-checked reporting
This is the kind of release that makes the open-weight delay sting a little more. A 600B-total model with a 1M context and sane API pricing is useful now; a locked box with a future open release is just a teaser trailer with a billing page. The real test is whether these giant agent models can stay practical without turning every task into a monologue.
Read more about this at: MarkTechPost
