---

license: apache-2.0

base_model: FINAL-Bench/Darwin-28B-Coder

base_model_relation: quantized

library_name: gguf

pipeline_tag: text-generation

tags:

• gguf
• llama.cpp
• mtp
• multi-token-prediction
• speculative-decoding
• code

language:

• en
• ko

---

> ### 📱 Run it on your phone or a GPU-less PC → POCKET · 🚀 Try it live (CPU chat)

> VIDRAFT's on-device family: a 35B model that runs on iPhone and on CPU with no GPU — stock llama.cpp, no fork.

>

>     

>

Darwin-28B-Coder — GGUF (MTP-enabled)

GGUF builds of FINAL-Bench/Darwin-28B-Coder with the native Multi-Token Prediction (MTP) head preserved, for self-speculative decoding in llama.cpp.

Requested in the base model discussion.

Files

| File | Quant | Size | Notes |

|------|-------|------|-------|

| Darwin-28B-Coder-Q4_K_M.gguf | Q4_K_M | 16.8 GB | recommended for most GPUs |

| Darwin-28B-Coder-Q8_0.gguf | Q8_0 | 29.0 GB | near-lossless |

| Darwin-28B-Coder-F16.gguf | F16 | 54.7 GB | full precision |

All files include the MTP layer — verified in metadata:

general.architecture = qwen35, qwen35.nextn_predict_layers = 1, tensors blk.64.nextn.*.

Multi-Token Prediction (MTP)

This model ships with a trained MTP head (1 prediction layer). With a recent llama.cpp build that includes MTP support (merged in PR #22673), the nextn layer is used for self-speculative decoding — typically ~1.5–2× faster generation with identical output (the main model verifies every drafted token, so quality is unchanged).

> A standard (non-MTP) GGUF does not contain the prediction head — you need these MTP-enabled files to benefit from the speedup.

Usage

# 1) Build a recent llama.cpp (MTP support is in mainline since PR #22673)
git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp
cmake -B build -DGGML_CUDA=ON && cmake --build build -j --config Release

# 2) Run — the nextn (MTP) layer enables self-speculative decoding
./build/bin/llama-cli \
  -m Darwin-28B-Coder-Q4_K_M.gguf \
  -ngl 99 -c 8192 \
  -p "Write a quicksort in Python."

For the exact MTP/speculative flags and the latest behaviour, see the llama.cpp MTP documentation / PR #22673. Works with llama-cli and llama-server.

Model spec (public)

| | |

|---|---|

| Architecture | qwen35 (hybrid attention) |

| Layers | 64 + 1 MTP |

| Hidden size | 5120 |

| Attention heads | 24 (KV 4) |

| Context length | 262,144 |

| Vocab | 248,320 |

| Precision (source) | bfloat16 |

License & attribution

License and usage follow the base model FINAL-Bench/Darwin-28B-Coder. These are GGUF conversions only; refer to the base model card for model details, intended use, and limitations.

GGUF conversion + quantization by the FINAL-Bench team using llama.cpp/convert_hf_to_gguf.py.

Learn more

• Running large coding models on-device without a GPU: Can you run a large LLM without a GPU?
• What model quantization is: What is model quantization?