---
license: apache-2.0
base_model:
- stepfun-ai/Step-3.7-Flash
library_name: gguf
language:
- en
tags:
- gguf
- llama.cpp
- mtp
- draft-model
- speculative-decoding
- q4_k_m
- q6_k
- q8_0
- bf16
- stepfun
---

# Step 3.7 Flash MTP Draft GGUFs

These are companion MTP draft GGUFs for speculative decoding. They are not standalone chat models.

Use them with the full model repo:

[notSnix/Step-3.7-Flash-Q4_K_M-GGUF](https://huggingface.co/notSnix/Step-3.7-Flash-Q4_K_M-GGUF)

The draft GGUF is passed with `--model-draft`; the full model is passed with `--model`.

## Files

| File | Size | SHA256 | Purpose |
|---|---:|---|---|
| `Step-3.7-Flash-MTP-Q8_0.gguf` | 3.5 GB | `017de8990140621b5b4af431448f20873fbf0b052f6c50d2afac15f45802a98d` | Recommended MTP draft |
| `Step-3.7-Flash-MTP-Q6_K.gguf` | 2.7 GB | `f41736e0dcce133d0dd0b81e14bd2965091e27dff306a28cec11ceb19fadbf46` | Smaller Q6_K MTP draft |
| `Step-3.7-Flash-MTP-Q4_K_M.gguf` | 2.0 GB | `44118cfe64f45b38127ad6fb626e16bd94ee5a827cb34aa83d9e6df3450aebaf` | Smaller MTP draft |
| `Step-3.7-Flash-MTP-BF16.gguf` | 6.5 GB | `fd811c81d14c786d314d8006655bba61971059abcfdfb6109ce83fd768f8b289` | Experimental BF16 MTP draft |

## Runtime

Current llama.cpp `main` supports Step MTP-tail draft loading natively. This was smoke-tested with clean llama.cpp commit `d545a2a993849fcf3b752d85ae256fc9d6a9de79`.

```bash
git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
cmake -B build-cuda -DGGML_CUDA=ON
cmake --build build-cuda --config Release -j
```

## Usage

```bash
llama-server \
  --model Step-3.7-Flash-Q4_K_M.gguf \
  --model-draft Step-3.7-Flash-MTP-Q8_0.gguf \
  --host 0.0.0.0 \
  --port 8000 \
  --ctx-size 262144 \
  --n-gpu-layers all \
  --split-mode layer \
  --parallel 1 \
  --reasoning on \
  --reasoning-format deepseek \
  --spec-type draft-mtp \
  --spec-draft-n-max 2 \
  --spec-draft-p-min 0.60
```

## Which Draft Should I Use?

Use `Step-3.7-Flash-MTP-Q8_0.gguf` first. It was the best local default in testing.

Use `Step-3.7-Flash-MTP-Q6_K.gguf` if you want a smaller draft file while staying above Q4.

Use `Step-3.7-Flash-MTP-Q4_K_M.gguf` if you want the smaller draft file.

Use `Step-3.7-Flash-MTP-BF16.gguf` for experimentation.

## Checksums

```bash
sha256sum -c SHA256SUMS
```

## Notes

- These files intentionally keep the upstream Step MTP tail-layer numbering (`blk.45`, `blk.46`, `blk.47`).
- They are companion speculative-decoding draft GGUFs, not full-model quants.
- The full Q4_K_M model is hosted separately so Hugging Face's GGUF widget does not display draft files as tiny full-model quantizations.
- This is a community GGUF conversion of the upstream Apache-2.0 model, not an official StepFun release.