Can I run step3 locally?
step3 by stepfun-ai has 321B parameters. At 4-bit with an 8K context it needs about 214 GB of GPU memory. It fits on 3 of 65 GPUs and Macs we track.
Memory needed by quant and context length
GB of GPU memory = weights + KV cache (FP16) + runtime overhead. Weights use parameter count × bits per weight. How we calculate.
| Quant | Quality | 4K ctx | 8K ctx | 32K ctx |
|---|---|---|---|---|
| 2-bit (Q2_K) | rough | 128 GB | 143 GB | 235 GB |
| 3-bit (Q3_K_M) | usable | 162 GB | 177 GB | 269 GB |
| 4-bit (Q4_K_M) | good, most popular | 199 GB | 214 GB | 305 GB |
| 5-bit (Q5_K_M) | very good | 229 GB | 244 GB | 336 GB |
| 6-bit (Q6_K) | near lossless | 261 GB | 276 GB | 368 GB |
| 8-bit (Q8_0) | effectively lossless | 334 GB | 349 GB | 440 GB |
| 16-bit (F16/BF16) | original weights | 614 GB | 629 GB | 720 GB |
Which GPUs and Macs run step3?
Best quality quant that fits at 8K context. "Tight" = within 10% of usable memory. Speed is a ceiling from memory bandwidth (real speed is typically 50–80% of it).
Questions
How much VRAM does step3 need?
About 214 GB at 4-bit (Q4_K_M) with an 8K context: 183 GB of weights (estimated), 30.5 GB of KV cache and 0.6 GB of runtime overhead. At 8-bit it needs about 349 GB.
What is the smallest GPU that runs step3?
No single consumer GPU in our table fits it at Q4_K_M with 8K context; the smallest option that does is the Mac M3 Ultra 512GB (512 GB).
Can I run step3 on a Mac?
Yes. At Q4_K_M with 8K context it fits on a Mac M3 Ultra 512GB or larger, using the default macOS GPU memory limit.
Model facts
| Hugging Face repo | stepfun-ai/step3 |
| Parameters | 320,969,912,832 (321B) |
| Layers / hidden size | 61 / 7168 |
| Attention | MHA; 64 heads, 64 KV heads, head dim 256 |
| KV cache per token (FP16) | 3904.0 KiB |
| Max context | 65,536 tokens |
| License | apache-2.0 |
| Released on HF | Jul 28, 2025 |
| Downloads (30d) / likes | 43,535 / 166 |
Sources: Hugging Face model API · config.json (stepfun-ai/step3) · GPU specs from vendor pages (list). Data fetched Nov 17, 2026. Spot an error? Report it.