base_model:
- hotdogs/Qwen3.8-27B-abliterated-code-analysis-preview
library_name: transformers
model_type: qwen3_5
pipeline_tag: text-generation
tags: - gguf
- qwen3
- code-review
- code-analysis
- lora
- sft
- abliterated
- multi-token-prediction
- mtp
- llama.cpp
license: mit
Qwen3.8-27B Code Analysis Preview (v2) — MTP GGUF
GGUF quantizations of the code-analysis fine-tune hotdogs/Qwen3.8-27B-abliterated-code-analysis-preview, with Multi-Token-Prediction (MTP) tensors preserved for speculative decoding.
Given a code snippet, it returns a structured, multi-paragraph review — real bugs, line-level reasoning, severity, and a concrete fix in a code block. It is a reasoning model: it thinks first, then answers.
v2 fixes the template-collapse of v1. v1 was trained on a synthetic placeholder dataset and answered in one line ("No bugs found. Code is clean."). v2 is retrained on 21,009 real code+bug+answer rows across 5 languages (Python, JS, Go, Rust, C) with detailed 550–880 char answers — the model now actually finds the bugs and generalizes to unseen bug types.
Files
| File | Size | Quant | Bits/weight |
|---|---|---|---|
Qwen3.8-27B-code-analysis-preview-v2-mtp-f16.gguf |
51 GB | F16 | 16.0 |
Qwen3.8-27B-code-analysis-preview-v2-mtp-Q6_K.gguf |
21 GB | Q6_K | 6.56 |
Qwen3.8-27B-code-analysis-preview-v2-mtp-Q4_K_M.gguf |
16 GB | Q4_K_M | 4.92 |
All 3 files: 866 tensors, MTP preserved — 15 blk.64.* tensors (11 transformer-layer + 4 blk.64.nextn.*).
MTP / speculative decoding
The MTP head lives in block 64 (blk.64.nextn.*). With llama.cpp you can use it as a draft model for speculative decoding:
llama-server -m Qwen3.8-27B-code-analysis-preview-v2-mtp-Q6_K.gguf \
--n-gpu-layers 999 --ctx-size 262144 --parallel 1 \
--cache-type-k f16 --cache-type-v f16 --flash-attn on \
--temp 1 --top-k 20 --top-p 0.95 --min-p 0.0 --jinja --tools all
(The blk.64.nextn.* tensors load automatically; enable speculative decoding via the predictor/draft options of your llama.cpp build.)
Smoke test (v2)
| Case | Result |
|---|---|
| Off-by-one (in-archetype) | 🟢 Found it + fix + docstring note |
| Async race (unseen) | 🟢 "no cache-hit fast path" + concurrency |
| Clean code (hallucination test) | 🟢 "correct, no bugs" + minor float/bool note |
Measured on Q6_K (5×3090 / 2 GPUs, flash-attn): ~34 t/s generation, ~141 t/s prompt eval.
Recommended
- Q6_K — best quality/size balance (21 GB) — recommended
- Q4_K_M — fastest / smallest (16 GB), great for consumer GPUs
- F16 — max fidelity (51 GB)
License
MIT (inherits the abliterated base).