Of the 600B, each step uses 27B
StepFun's Step 5 Preview card on Hugging Face describes a sparse mixture-of-experts model with 600B total parameters and 27B active per token. Context is up to 1 million tokens, and maximum output is also listed as up to 1 million tokens. Inputs include text, images, and video. Output is text.
On the same public record, the repository was created at 03:15 UTC on September 20 and last modified at 01:28 UTC on October 15, which is 09:28 Beijing time today. The creation day and today's modification are not the same day. This piece does not treat today's update as the day the model first appeared.
[1]
One million tokens is on the table, and memory is not guaranteed
The card gives deployment examples: vLLM or SGLang behind a compatible API, and Transformers for offline generation, in FP8 or BF16. The example machines include four or eight B300 GPUs, and eight or sixteen H200 GPUs. Those are the starting configurations they wrote down.
The card then says these configurations do not guarantee enough memory for a 1-million-token request. The KV cache and runtime buffers also take GPU memory. Usable context and concurrency vary with the checkpoint, the inference framework, and cache allocation.
[1]The two-node BF16 shard path is not fully checked
FP8 and BF16 need their own checkpoints. The card says setting the dtype on an FP8 checkpoint does not turn its quantized weights into BF16.
The two-node, 16-GPU BF16 path first splits the checkpoint into 16 shards and then generates. The card says this full path, from conversion through generation by shard, has not yet been validated end to end. What has been checked is checkpoint names and shapes, BF16 conversion and generation on a small model, and full-model streaming-weight inference.
[1]The license adds conditions beyond the familiar grant
The card says the weights, parameters, configuration, tokenizer, model-specific code, and documentation are under the StepFun Model License 1.0. It calls that license Apache-2.0-style, with additional attribution and commercial-license conditions. The full text is the LICENSE file in the repository.
The card also points part of the evaluation discussion at an external snapshot dated September 20, and says the pictured results use the High reasoning setting. The extracted text has no score table that can be copied, so this piece does not add scores.
[1]