The Pulse
StepFun Ships 600B Step 5 Preview With 1M-Token Context
StepFun has released Step 5 Preview, a sparse-MoE model with 600 billion total parameters, 27 billion activated parameters, a 1-million-token context window and fully open API access. The source contains no attributable statement from a nam
AI.info Team ·
StepFun has released a model that puts two competing claims on the same specification sheet: a total parameter count of 600 billion and an inference footprint of 27 billion active parameters. Step 5 Preview also supports a 1-million-token context window, giving the Chinese AI company a flagship aimed at long-running software, research, and financial tasks rather than short chat exchanges.
Step 5 Preview was announced on September 20. Full API access is open now, while the company plans to release the model weights on October 15. 凤凰网科技 reported the launch and specifications.
600B Parameters, 27B Active Per Token
Step 5 Preview uses a sparse mixture-of-experts architecture. Its 600 billion parameters represent the model’s total capacity, while only 27 billion parameters are activated for each token during inference. That separation allows StepFun to offer a model with a large overall capacity without activating the full parameter count for every token.
The architecture accepts text and visual inputs and produces text outputs. StepFun says the model is designed for real-world agentic tasks, including software engineering, financial research, front-end design, 3D asset generation, and other workflows that require extended execution rather than a single response.
The Million-Token Window Is the Main Bet
The 1-million-token context window is the model’s most consequential user-facing specification. It allows a single workflow to retain far more source material, code, tool results, and intermediate information than models limited to conventional context lengths.
StepFun presents the model as suited to long-cycle tasks. In one reported example, Step 5 Preview worked on GPU-kernel optimization for 24 hours and raised peak MLA-kernel performance to 508 TFLOPS, compared with 493 TFLOPS for Claude Opus 5. The report also says an automated post-training experiment raised Qwen3-30B-A3B’s AIME24 accuracy from 53.3% to 60%.
Step 5 Preview Lands Near the Top of Artificial Analysis
Step 5 Preview scores 44 on the Artificial Analysis Intelligence Index, placing it among the top three open models on that ranking. The report also says the model’s cost per completed task is one-eighth that of Anthropic’s Claude Opus 5.
Those figures describe a cost-and-capability comparison rather than a universal performance win. Benchmark results can vary with model settings, prompting, tool access, and evaluation design. The report says Step 5 Preview performed just behind GPT-6 Astra or Claude Opus 5 on the CLI subset of Agents’ Last Exam, the FrontierFinance financial-investment research evaluation, and the DRACO cross-domain deep-research evaluation, while leading other participating open models.
Software Engineering Comes Before the Weight Release
StepFun’s self-built StepCodeBench covers 553 code repositories, nine task types, and 33 programming languages. The report says the model showed strong results in overall task success rate and stability across scenarios.
The report also describes application scenarios involving modification of an ESP32 development board, front-end design, 3D asset generation, and the full financial-research workflow. It says the ESP32 task continued for more than three hours.
API access is available now, while the full model weights are scheduled for release on October 15. That release will provide a broader test of whether a 600-billion-parameter model with 27 billion active parameters can combine long-context persistence, multimodal input, and open availability without imposing deployment costs that erase its efficiency claim.