JuicyTalk
StepFun has released Step 5 Preview, its new flagship model for agentic work. The target workloads are software engineering, professional knowledge work, and finance. The main pitch is cost. StepFun team states the model delivers comparable intelligence at a substantially lower task cost. That is the ‘Pareto frontier’ framing in the launch title.
Is it deployable? Yes, as a hosted API and on the StepFun platform. Self-hosting waits for open weights. StepFun says open weights land on October 15, 2026. By simple arithmetic, 600B parameters need about 1.2 TB in BF16, before KV cache. Plan for multi-GPU server hardware once weights ship.
What StepFun Shipped
Step 5 Preview is a sparse Mixture-of-Experts (MoE) model. It holds about 600B total parameters and activates about 27B per token. That is roughly 4.5% of the weights per token.
The official model documentation lists these specs:
- Model ID:
step-5-preview - Context window: 1M tokens
- Input: text, images, and video
- Output: text
- Reasoning effort:
low,medium, andhigh - Streaming, tool calling, JSON Mode, JSON Schema, and prompt caching
On research tasks, StepFun team states the model coordinated 950 web fetches in a single agent action. StepFun team also documents a Claude Code integration through its Step Plan.
Architecture: Narrow and Deep
StepFun did not widen the network. It stacked 92 Transformer layers in a narrow-deep layout, according to Pandaily. The research team argues deeper stacks give longer paths for implicit multi-hop reasoning. This matters during long prefill, when agents search, run code, and read tool returns.
Training leans on on-policy, long-horizon reinforcement learning. StepFun cites bit-wise train and inference alignment across MoE routing. Other listed techniques include MTP-3 speculative decoding, FP8 MoE, and KV-cache offload. StepFun reports more than 3x end-to-end speedup for long-horizon RL.
