- Published on
StepFun Launches Step 5 Preview: A 600B-Total, 27B-Active MoE Model With 1M Context for Long-Horizon Agentic Work
- Authors
- Name
- DeepHorus
StepFun has released Step 5 Preview, a sparse Mixture-of-Experts flagship for agentic software engineering, professional knowledge work, and finance. The model has about 600B total parameters, activates 27B per token, and supports a 1M-token context with text, image, and video input. StepFun pitches it as a cost frontier, delivering comparable intelligence at lower task cost.
The architecture uses 92 Transformer layers in a narrow-deep layout for implicit multi-hop reasoning. Training relied on long-horizon reinforcement learning, FP8 MoE, MTP-3 speculative decoding, and KV-cache offload. StepFun reports over 3x end-to-end speedup for long-horizon RL.
The model is available through a hosted API. Pricing is $1.00 per million input tokens and $2.70 per million output tokens. Artificial Analysis puts comparable reasoning-model medians at $1.88 and $10.00, but notes Step 5 Preview generated 160M output tokens versus a 92M median, so verbose reasoning reduces savings.
On company benchmarks at High effort, Step 5 Preview scored 67.7 on DeepSWE v1.1, 49.0 on StepCodeBench, and 80.5 on ProgramBench. GPT-6 Astra and Claude Opus 5 remain ahead on all three. In 24-hour agent runs, StepFun says it tuned an H100 kernel to 508 TFLOPS and raised Qwen3-30B-A3B on AIME24 from 53.3% to 60%.
Artificial Analysis scored it 44 on its Intelligence Index, versus a median of 24 for similar-priced reasoning models, and measured 99.8 output tokens per second. The API supports streaming, tool calling, JSON Mode, JSON Schema, and prompt caching.
Open weights are scheduled for October 15, 2026. A BF16 deployment of 600B parameters would need about 1.2 TB before KV cache, so multi-GPU hardware will be required. Until then, access is limited to the hosted API.
The release intensifies competition in cost-efficient agentic AI, though independent verification remains limited. StepCodeBench is StepFun's own benchmark, and its coding results still trail leading rivals. The open-weight release will let the community test the architecture and cost claims directly.
