A legal-domain fine-tune of huihui-ai/Huihui-Qwen3.6-27B-abliterated (an abliterated Qwen3.6-27B vision-language model), trained on 21k+ public-domain United States court opinions from CourtListener.

All base capabilities are preserved and were verified after merging: vision, tool/function calling, and thinking mode.

What it is good at

Legal reasoning, doctrine, and terminology (constitutional criminal procedure is well represented)
Drafting in judicial-opinion and legal-memo style, including parenthetical case summaries
Fresh coverage of 2025-2026 Fourth Amendment location-data rulings, including Chatrie v. United States (S. Ct., June 29, 2026, geofence warrants), which post-date most training corpora

IMPORTANT: pair it with retrieval

Fine-tuning teaches doctrine and style, not verbatim recall. Like every LLM, this model can hallucinate reporter citations, dates, and quotes. For any real legal-research use, run it with RAG over the CourtListener bulk data (or the linked training corpus) so citations come from retrieved documents. Verify every citation at the source before relying on it.

Training

Method: QLoRA (r=64, q/k/v/o/gate/up/down) via Unsloth on 1x A100-80GB; LoRA merged into the pristine bf16 base afterward, which is why vision/tools/thinking survive intact
Data: 16,000 CPT opinion documents + 12,000 SFT holding-summary pairs + 41 recent rulings, 2,048-token sequences, 1 epoch (~13 h)
Final train loss: 1.385

Variants

Repo	Format	Size	For
this repo	bf16 safetensors	52 GB	further fine-tuning, serving on 80GB+
-FP8	FP8 W8A8 (compressed-tensors)	29 GB	vLLM on 40GB+ GPUs
-GGUF	Q4_K_M GGUF + vision projector	16 GB	Ollama / llama.cpp on 24-32GB GPUs

Quick start (Ollama)

Download the GGUF variant files, then:

# Modelfile
FROM ./legal-27b.Q4_K_M.gguf
FROM ./legal-27b-mmproj.gguf
RENDERER qwen3.5
PARSER qwen3.5

ollama create legal-27b -f Modelfile
ollama run legal-27b

Quick start (vLLM, FP8 variant)

bash

vllm serve scottyjmp5/Legal-Qwen3.6-27B-Abliterated-FP8 \
  --trust-remote-code --max-model-len 8192

Warnings

Abliterated base: refusal behaviors of the original Qwen3.6 have been removed upstream. You are responsible for output filtering appropriate to your deployment.
Not legal advice. Outputs are drafts/research aids and must be reviewed by a licensed attorney.
US-centric: training data is exclusively United States case law.