Can I run Step 3.5 Flash locally?
Step 3.5 Flash by stepfun-ai has 199B parameters. At 4-bit with an 8K context it needs about 116 GB of GPU memory. It fits on 13 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 real file sizes from bartowski/stepfun-ai_Step-3.5-Flash-GGUF where available (marked •), otherwise parameter count × bits per weight. How we calculate.
| Quant | Quality | 4K ctx | 8K ctx | 32K ctx | 128K ctx |
|---|---|---|---|---|---|
| 2-bit (Q2_K) • | rough | 67.5 GB | 69.0 GB | 78.0 GB | 114 GB |
| 3-bit (Q3_K_M) • | usable | 86.7 GB | 88.2 GB | 97.2 GB | 133 GB |
| 4-bit (Q4_K_M) • | good, most popular | 114 GB | 116 GB | 125 GB | 161 GB |
| 5-bit (Q5_K_M) • | very good | 133 GB | 135 GB | 144 GB | 180 GB |
| 6-bit (Q6_K) • | near lossless | 154 GB | 155 GB | 164 GB | 200 GB |
| 8-bit (Q8_0) • | effectively lossless | 198 GB | 199 GB | 208 GB | 244 GB |
| 16-bit (F16/BF16) | original weights | 374 GB | 376 GB | 385 GB | 421 GB |
Which GPUs and Macs run Step 3.5 Flash?
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 Step 3.5 Flash need?
About 116 GB at 4-bit (Q4_K_M) with an 8K context: 112 GB of weights (real GGUF file size), 3.6 GB of KV cache and 0.6 GB of runtime overhead. At 8-bit it needs about 199 GB.
What is the smallest GPU that runs Step 3.5 Flash?
No single consumer GPU in our table fits it at Q4_K_M with 8K context; the smallest option that does is the H200 (141 GB).
Can I run Step 3.5 Flash on a Mac?
Yes. At Q4_K_M with 8K context it fits on a Mac M2 Ultra 192GB or larger, using the default macOS GPU memory limit.
Model facts
| Hugging Face repo | stepfun-ai/Step-3.5-Flash |
| Parameters | 199,384,301,376 (199B) |
| Layers / hidden size | 45 / 4096 |
| Attention | MHA + sliding window; 64 heads, 64 KV heads, head dim 128; sliding window 512 |
| KV cache per token (FP16) | 1536.0 KiB (sliding layers capped at the window) |
| Max context | 262,144 tokens |
| License | apache-2.0 |
| Released on HF | Feb 1, 2026 |
| Downloads (30d) / likes | 180,796 / 847 |
Sources: Hugging Face model API · config.json (stepfun-ai/Step-3.5-Flash) · GGUF files · GPU specs from vendor pages (list). Data fetched Nov 17, 2026. Spot an error? Report it.