StepFun launches Step 5 Preview, a 600B-parameter MoE model

StepFun introduced Step 5 Preview, a new flagship model built on a sparse Mixture-of-Experts architecture with 600B total parameters and 27B active per token, supporting a 1M-token context window and vision input. StepFun positions it as an agentic model for software engineering, professional knowledge work and finance, and frames the release around the Pareto frontier of intelligence versus cost: Step 5 Preview scores 44 on the Artificial Analysis Intelligence Index, and StepFun says that at a comparable level of intelligence its task cost is substantially lower than similarly capable models. It is available today through StepFun's products and API, with open weights planned for October 15 (the announcement does not state a year for that date).

On coding, StepFun built an internal benchmark, StepCodeBench, spanning 553 independent repositories, 9 task categories, 20 application domains and 33 programming languages; Step 5 Preview scores an avg@4 of 49.0% on it, performing particularly well on bug repair, feature modification and refactoring, and beating open-weight peers GLM-5.3 (40.2%) and Kimi K3 (43.9%) while trailing Claude Opus 5 (63.9%) and GPT-6 Astra (61.0%). Across internal and external expert evaluations, about 70% of participants judged Step 5 Preview capable of autonomously solving coding tasks of moderately high complexity. On a broader published benchmark table, it also trails Claude Opus 5 and GPT-6 Astra on DeepSWE v1.1 (67.7% versus 74.0% and 74.1%) and Terminal-Bench v4 (33.3% versus 52.3% and 57.9%), ties Kimi K3 on GPQA Diamond at 93.5% (below GPT-6 Astra's 96.1%), and scores 46.5% on HLE, below Claude Opus 5's 54.9%. Given development documentation and user authorization, StepFun says the model can also work directly with programmable hardware, using cameras, COM ports, screenshots and simulated mouse input during development and debugging.

StepFun also tested long-horizon execution. Given 24 hours and four independent attempts to optimize an MLA GPU kernel from scratch on an NVIDIA H100, Step 5 Preview reached a peak of 508 TFLOPS for forward and backward passes after roughly 22 hours, ahead of Claude Opus 5's 493 TFLOPS on the same task. In a second test, it had 24 hours to improve a Qwen3-30B-A3B base model on AIME24 through automated post-training with an API annotator; the resulting model reached 60% accuracy, up from 53.3% before post-training, matching the result Claude Opus 5 achieved while using fewer annotator tokens. As a further, non-coding stress test, StepFun ran Step 5 Preview on Pokemon Red without any game-specific optimization: the model sustained progress for more than 3,000 turns and 6 million tokens of interaction, and by turn 3,082 had unlocked the move Cut, earned three Gym Badges and defeated Lt. Surge, putting the run roughly a third of the way through a main story that takes an estimated 26 hours for a human to complete.

On professional knowledge work, StepFun describes a climate study covering 1,000 locations over 25 years, in which Step 5 Preview coordinated 950 web fetches in a single agent action and assembled 300,000 monthly records across 11 variables, using the dataset to find that peak solar-seasonality months fell at opposite times of year in the European and Oceanian samples. In a diesel surcharge review, it produced a 17-sheet analytical workbook containing source data, cross-series reconciliation, regional panels, formulas and trend models, and in a separate task it produced an interactive report combining written analysis, visualizations, data tables and methodological notes.

Finance was a specific focus: StepFun built three internal benchmarks, FinStepBench-LiveSearch (retrieving and verifying timely financial information), FinStepBench-CorporateValuation (turning financial data into consistent, reproducible valuations) and FinStepBench-DeepResearch (the full research process from evidence gathering to a supported report), and says the work behind financial analysis must be open to scrutiny: a reviewer should be able to trace a conclusion back to its sources, distinguish reported facts from assumptions, and reproduce the calculations. Step 5 Preview was also evaluated on FrontierFinance, an external benchmark covering six investment use cases through 220 expert-crafted questions and 11,543 evaluation criteria, where it scored 66.4, ahead of GLM-5.3 (64.1), Kimi K3 (62.6) and GPT-6 Astra (55.0) but behind Claude Opus 5 (69.7).

Key facts

Step 5 Preview is a sparse MoE model with 600B total parameters, 27B active per token, a 1M-token context window and vision input; it scores 44 on the Artificial Analysis Intelligence Index.
On a 24-hour MLA GPU kernel optimization test, it reached a peak of 508 TFLOPS after roughly 22 hours, ahead of Claude Opus 5's 493 TFLOPS on the same task.
In 24 hours of automated post-training, it lifted a Qwen3-30B-A3B model's AIME24 score from 53.3% to 60%, matching Claude Opus 5's result with fewer annotator tokens.
Without game-specific tuning, it sustained a Pokemon Red run past 3,000 turns and 6 million tokens, reaching three Gym Badges and about a third of the main story.
It scored 66.4 on the external FrontierFinance benchmark, ahead of GLM-5.3, Kimi K3 and GPT-6 Astra but behind Claude Opus 5's 69.7; open weights are planned for October 15.

Why it matters

StepFun frames Step 5 Preview around the Pareto frontier of intelligence versus cost: a 44 on the Artificial Analysis Intelligence Index paired with, StepFun says, substantially lower task cost than similarly capable models at comparable intelligence. The architecture (600B total parameters, 27B active per token, sparse Mixture-of-Experts, 1M-token context) is built for agentic, long-horizon work rather than single-turn question answering. The GPU kernel optimization, the post-training experiment and the Pokemon Red run are all designed to show sustained execution across many steps, not just a benchmark score.

Who it affects

Teams building agentic coding tools now have another frontier-adjacent option to weigh against Claude Opus 5, GPT-6 Astra, GLM-5.3 and Kimi K3. Financial analysis and due-diligence teams are a specific target, given the three purpose-built FinStepBench evaluations and the FrontierFinance results. The open-weight community is affected directly once weights ship on October 15, since that lets anyone self-host the model.

How to use it

Step 5 Preview is available today through StepFun's own products and API; the announcement gives no pricing or subscription terms. Open weights are planned for October 15, though the announcement does not state which year, so self-hosting is not yet possible from this release.

How solid is it

The headline numbers rest on StepFun's own internal benchmarks (StepCodeBench, the FinStepBench suite) plus a couple of external ones (the Artificial Analysis Intelligence Index, GDPval-AA v2, FrontierFinance). The GPU kernel and post-training comparisons against Claude Opus 5 are StepFun's own 24-hour internal test setups, not independently audited runs. On the full published benchmark table, Step 5 Preview beats open-weight peers GLM-5.3 and Kimi K3 on several metrics but trails Claude Opus 5 and GPT-6 Astra on many coding and agentic ones, including DeepSWE v1.1, Terminal-Bench v4, StepCodeBench and GDPval-AA v2, while leading on a smaller set such as AA-LCR v1.1.

Risks and caveats

The announcement does not explain why Step 5 Preview trails Claude Opus 5 and GPT-6 Astra on several benchmarks (StepCodeBench, Terminal-Bench v4, HLE) while leading on others. No API pricing or subscription cost is given anywhere in the source. No named individual is credited; every claim is made in a collective "we". The open-weight release date, October 15, appears without a stated year in that sentence.

“A reviewer should be able to trace a conclusion back to its sources, distinguish reported facts from assumptions, and reproduce the calculations.”

— StepFun, Step 5 Preview announcement