---

base_model: deepseek-ai/deepseek-coder-6.7b-base

tags:

• gguf
• fim
• fill-in-the-middle
• code
• rust
• reth
• llama.cpp
• ollama

pipeline_tag: text-generation

license: other

license_name: deepseek

license_link: https://github.com/deepseek-ai/DeepSeek-Coder/blob/main/LICENSE-MODEL

---

deepseek-coder-6.7b-fim-reth-v1 — GGUF

GGUF quantizations of the deepseek-coder-6.7b-fim-reth-v1

LoRA adapter merged into DeepSeek-Coder-6.7B base.

Fine-tuned for Fill-in-the-Middle (FIM) code completion on the reth Rust codebase.

Files

| File | Quant | Size | Notes |

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

| fim-deepseek-6.7b-reth-v1.q4_k_m.gguf | Q4_K_M | 4.08 GB | Fast, small — good default for laptops / Ollama |

| fim-deepseek-6.7b-reth-v1.q8_0.gguf | Q8_0 | 7.16 GB | Near-lossless quality |

FIM prompt format

DeepSeek FIM tokens — completion is generated after <｜fim▁end｜>:

<｜fim▁begin｜>{prefix}<｜fim▁hole｜>{suffix}<｜fim▁end｜>

Usage

llama.cpp

llama-cli -m fim-deepseek-6.7b-reth-v1.q4_k_m.gguf \
  -p $'<｜fim▁begin｜>fn add(a: i32, b: i32) -> i32 {\n    <｜fim▁hole｜>\n}<｜fim▁end｜>' \
  -n 64 --temp 0.2 -no-cnv

Ollama

FROM ./fim-deepseek-6.7b-reth-v1.q4_k_m.gguf
TEMPLATE """{{ .Prompt }}"""
PARAMETER temperature 0.2
PARAMETER num_predict 128
PARAMETER num_ctx 4096
PARAMETER stop "<｜fim▁begin｜>"
PARAMETER stop "<｜fim▁hole｜>"
PARAMETER stop "<｜fim▁end｜>"
PARAMETER stop "<｜end▁of▁sentence｜>"

ollama create fim-deepseek-reth -f Modelfile

Evaluation (parent adapter, 200 held-out reth samples)

| Metric | Tuned | Base | Δ |

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

| pass@1 (exact match) | 31.0% | 13.5% | +17.5 pts (2.3×) |

| Edit similarity | 0.650 | 0.453 | +0.197 |

| BLEU | 0.459 | 0.294 | +0.165 |

GGUF-verified on real held-out reth code

These GGUF files were independently re-benchmarked via llama.cpp (greedy, temp 0) on

200 real held-out reth FIM samples — generated completion vs. the actual reth code:

| Model | pass@1 | edit-sim | size | speed (M4 Pro) |

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

| q8_0 (tuned) | 18.0% | 0.575 | 6.7 GB | ~23 tok/s |

| q4_k_m (tuned) | 16.0% | 0.559 | 3.8 GB | ~50 tok/s |

| base (q8_0) | 7.5% | 0.442 | — | — |

→ tuned is ~2.4× (q8) / ~2.1× (q4) the base model's pass@1 on reth — reproducing the

adapter's reported lift end-to-end in GGUF. q4_k_m keeps ~90% of the gain at half the

size; recommended default for laptops.

License

Governed by the DeepSeek License Agreement for the base model.