Models / Step 5 Preview

Best home rig for Step 5 Preview

Current model600B total · 27B active1 measured result model info ↗

Will it fit my GPU? Best value GPU per euro →

Real, measured generation speeds for Step 5 Preview on home hardware, fastest first. Different quants/engines aren't directly comparable.

Build	Quant	Engine	tok/s	Cost	Source
4x DGX Spark (GB10) running Step 5 Preview with tensor parallel over RoCE	EXL3 (experts 4.01 bpw, 341.4 GB total)	vLLM, TP=4 over RoCE, MTP (2 draft tokens)	25 ⚡		github.com ↗

Rent before you buy: try this model on a cloud GPU for a few dollars first: Runpod (referral link; sign up with Google to get the new-user bonus)

Memory math: how much VRAM / RAM does Step 5 Preview need?

Step 5 Preview has 600B parameters, about 27B active per token (source: StepFun model card: 600B total, 27B active (MoE)). Weight size = parameters × bits-per-weight ÷ 8, before context (KV cache) and runtime overhead.

Format	Bits/weight	Weights	Smallest common memory size that holds the weights alone
Q8_0	8.50	637.6 GB	1024 GB
Q6_K	6.56	492.2 GB	512 GB
Q5_K_M	5.70	427.8 GB	512 GB
Q4_K_M	4.89	367.1 GB	512 GB
Q3_K_M	4.00	299.7 GB	512 GB
Q2_K	3.16	236.9 GB	256 GB
As measured: EXL3 (experts 4.01 bpw, 341.4 GB total)	4.01	≈300.8 GB	512 GB

Bits-per-weight for GGUF quants are llama.cpp's own figures (llama.cpp quantize README). "As measured" rows apply the source's stated bits-per-weight to the parameter count; real files differ slightly because some layers stay at higher precision. Try other combinations in the VRAM calculator.

Why it can still run on smaller GPUs: it is a mixture-of-experts model, so only about 27B of its 600B parameters are used for each generated token. Runtimes such as llama.cpp (--n-cpu-moe), ExLlamaV3 or Project Maya can keep most expert weights in system RAM and only the shared layers and hot experts in VRAM. Then you need VRAM + system RAM ≥ weights, and speed is limited mainly by RAM bandwidth.

FAQ

How much VRAM does Step 5 Preview need?

The weights are about 367.1 GB at Q4_K_M (600B params × 4.89 bits ÷ 8, before context/KV cache). Add memory for context (KV cache) and the runtime. The smallest-memory setup with a measured result here, fully in memory, is 4x DGX Spark (GB10) (512 GB) at 25 tok/s.

What is the fastest measured hardware for Step 5 Preview?

4x DGX Spark (GB10): 25 tok/s generation (vLLM, TP=4 over RoCE, MTP (2 draft tokens), EXL3 (experts 4.01 bpw, 341.4 GB total)), from https://github.com/0xSero/Step-5-Preview-Four-Sparks.

Which GPUs have measured results for Step 5 Preview?

4x DGX Spark (GB10).

Can it run on… 4x DGX Spark (GB10)