Evidence: benchmark results vendor-reported
Step 5 Preview 600B-A27B — VRAM, speed & local setup
Written by Jakub Rusinowski · Last updated
StepFun's 600B-parameter MoE with 27B active per token and up to 1M tokens of context, taking text, image and video input. 92 layers, 23 of them full attention and 69 with a 512-token sliding window. Weights in FP8, BF16 and NVFP4 since 15 October 2026. A server model: StepFun's examples serve it on multi-GPU nodes.
Step 5 Preview 600B-A27B needs about 366 GB of VRAM at Q4_K_M — quantized weights plus framework overhead, before any KV cache. On Apple Silicon that figure comes out of unified memory.
VRAM and speed by quantization
Quoted against NVIDIA RTX 4090 (24 GB). Includes the KV cache at 8K context, so it reads higher than the headline figure.
| Quant | Memory | VRAM | Speed (est.) | Fit |
|---|---|---|---|---|
| Q2_K 2.63 bpw | 200.3 GB | — | Too big | |
| Q3_K_M 3.41 bpw | 259.2 GB | — | Too big | |
| Q4_K_M 4.83 bpw | 366.4 GB | — | Too big | |
| Q5_K_M 5.67 bpw | 429.9 GB | — | Too big | |
| Q6_K 6.56 bpw | 497.1 GB | — | Too big | |
| Q8_0 8.50 bpw | 643.7 GB | — | Too big | |
| F16 16.00 bpw | 1210.2 GB | — | Too big |
Black marker = usable memory on the NVIDIA RTX 4090 (24 GB). Estimates from the memory-bandwidth roofline on the methodology page. Step 5 Preview 600B-A27B VRAM calculator →
Get Step 5 Preview 600B-A27B running
Step 5 Preview 600B-A27B is not practical on a single consumer GPU: it needs about 365.6 GB at Q4_K_M, more than any single device in our catalogue holds with headroom. How multi-GPU inference works, or rent a multi-GPU server by the hour:
As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.
How to run Step 5 Preview 600B-A27B
No Ollama tag and no official GGUF. StepFun documents vLLM and SGLang serving.
Specifications
Verified — Checked against the primary source — the model card or the vendor spec page — and corroborated by a second independent source.
- Parameters
- 600 Billion (27B active)
- Context window
- 1,048,576
- Architecture
- Mixture-of-Experts (352 routed + 1 shared, 8 active), GQA with 3:1 sliding-window to full attention
- Provider
- StepFun
- Licence
- StepFun Model License 1.0
- Practical deployment
- Multi-GPU server
- Specified at
- Q4_K_M
- System RAM
- 512 GB
- Record updated
- 2026-11-03
Weights are downloadable and commercial use is permitted, subject to the licence’s acceptable-use terms.
Above 100M monthly active users or $20M monthly revenue a product must display "Powered by StepFun AI"; device makers, and model-as-a-service operators above $20M revenue in any 12 months, need a separate commercial licence.
Limits and caveats
- A preview release by name; StepFun may change it.
- No official GGUF; the published formats are FP8, BF16 and NVFP4.
- Benchmarks on the card are vendor-reported.
Quality and use cases
Scores as published by the model’s authors or an independent evaluator — quality, not throughput, and not measured by us.
Other StepFun Step sizes
Step 5 Preview 600B-A27B — frequently asked questions
How much VRAM does Step 5 Preview 600B-A27B need?
About 366 GB at Q4_K_M — quantized weights plus framework overhead, before any KV cache. The cache grows with context length and is added on top; the table above folds it in. Apple Silicon counts unified memory toward the same figure.
Does Step 5 Preview 600B-A27B run on an RTX 4090 (24 GB)?
No. Step 5 Preview 600B-A27B needs about 366 GB at Q4_K_M, more than a single RTX 4090's 24 GB. It needs a larger card, several GPUs, or Apple Silicon with enough unified memory — or it runs with part of the weights offloaded to system RAM, which is much slower.
How do I run Step 5 Preview 600B-A27B locally?
No Ollama tag and no official GGUF. StepFun documents vLLM and SGLang serving. Running the published tag would send your prompts to a hosted GPU rather than your own machine.
What other sizes does StepFun Step come in?
Step 3.7 Flash 198B-A11B (112 GB), Step 5 Preview 600B-A27B (366 GB). Every size shares the family's training and licence; the larger ones score higher and need proportionally more memory.