StepFun released Step 5 Preview. Weights are on Hugging Face and ModelScope. It is a 600 billion parameter MoE. 27 billion parameters are active per token. Context is up to one million tokens.
What shipped
StepFun is a Chinese AI lab. It released Step 5 Preview. The model is a sparse Mixture-of-Experts.
The model has 600B total parameters. 27B are active per token. Context is up to 1M tokens. Maximum output is up to 1M tokens.
License is StepFun Model License 1.0. Hugging Face lists it as other. 2
Weights are public. Hugging Face FP8 is stepfun-ai/Step-5-Preview. 2 Hugging Face BF16 is stepfun-ai/Step-5-Preview-BF16. 4 ModelScope FP8 is stepfun-ai/Step-5-Preview. 5 ModelScope BF16 is stepfun-ai/Step-5-Preview-BF16. 6
FP8 checkpoint is about 650 GB. API launched September 20, 2026. Weights were released October 14, 2026. 1
Input is text and image. The API adds video.
What changed
API went live September 20, 2026. Open weights followed October 14, 2026.
The model uses 92 transformer blocks. StepFun describes it as narrow but deep. 3
Attention is Gated Sliding Window Attention plus Sparse GQA. History per query is 4,096 tokens. Block size is 8.
Speculative decoding uses MTP-3. StepFun claims about 30 percent end to end speedup. That is vendor claimed.
Training used FP8 mixed precision with FP32 hidden states.
BF16 Transformers workflow is not yet validated end to end, per StepFun.
Benchmarks from StepFun use High reasoning. Rivals in comparison charts ran at Max. That is not like for like.
Why it matters
This is datacenter hardware. It is not a home lab model.
FP8 on vLLM or SGLang needs 4x B300 or 8x H200 on one node.
BF16 on vLLM needs 8x B300 or 16x H200 across one to two nodes.
Transformers offline generation needs 8x H200 for FP8 or 16x H200 for BF16.
FP8 checkpoint is about 650 GB.
License is StepFun Model License 1.0. It is not Apache 2.0 or MIT. Read it before local deploy.
API pricing is $1.00 per 1M input tokens, $2.70 per 1M output tokens, $0.10 per 1M cache hits.
Poor fit for single GPU boxes or a 5090 or a DGX Spark. Good fit for labs with multi H200 or B300 nodes that want a 1M context open weights MoE with 27B active.
