StepFun has launched Step 5 Preview, its new flagship mannequin for agentic work. The goal workloads are software program engineering, skilled data work, and finance. The principle pitch is value. StepFun crew states the mannequin delivers comparable intelligence at a considerably decrease job value. That’s the ‘Pareto frontier’ framing within the launch title.
Is it deployable? Sure, 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 easy arithmetic, 600B parameters want about 1.2 TB in BF16, earlier than KV cache. Plan for multi-GPU server {hardware} as soon as weights ship.
What StepFun Shipped
Step 5 Preview is a sparse Combination-of-Specialists (MoE) mannequin. It holds about 600B whole parameters and prompts about 27B per token. That’s roughly 4.5% of the weights per token.
The official model documentation lists these specs:
- Mannequin ID:
step-5-preview - Context window: 1M tokens
- Enter: textual content, photographs, and video
- Output: textual content
- Reasoning effort:
low,medium, andexcessive - Streaming, instrument calling, JSON Mode, JSON Schema, and immediate caching
On analysis duties, StepFun crew states the mannequin coordinated 950 internet fetches in a single agent motion. StepFun crew additionally paperwork a Claude Code integration by way of its Step Plan.
Structure: Slender and Deep
StepFun didn’t widen the community. It stacked 92 Transformer layers in a narrow-deep structure, according to Pandaily. The analysis crew argues deeper stacks give longer paths for implicit multi-hop reasoning. This issues throughout lengthy prefill, when brokers search, run code, and skim instrument returns.
Coaching leans on on-policy, long-horizon reinforcement studying. StepFun cites bit-wise prepare and inference alignment throughout MoE routing. Different listed strategies embrace MTP-3 speculative decoding, FP8 MoE, and KV-cache offload. StepFun stories greater than 3x end-to-end speedup for long-horizon RL.
