license: apache-2.0
base_model: scottyjmp5/Legal-Qwen3.6-27B-Abliterated
datasets:
- scottyjmp5/courtlistener-legal-corpus
language: - en
pipeline_tag: image-text-to-text
tags: - legal
- caselaw
- abliterated
- vision
- tool-calling
- gguf
- ollama
- llama.cpp
Legal-Qwen3.6-27B-Abliterated-GGUF
GGUF build of Legal-Qwen3.6-27B-Abliterated
for Ollama / llama.cpp.
| File | Size | Notes |
|---|---|---|
legal-27b.Q4_K_M.gguf |
16 GB | text model, Q4_K_M |
legal-27b-mmproj.gguf |
0.9 GB | vision projector (required for image input) |
Runs fully on a single 24-32 GB GPU. Verified on an RTX 5090 at ~71 tok/s with vision,
tool calling, and thinking all working through Ollama 0.32+.
Ollama
# Modelfile
FROM ./legal-27b.Q4_K_M.gguf
FROM ./legal-27b-mmproj.gguf
RENDERER qwen3.5
PARSER qwen3.5
PARAMETER temperature 1
PARAMETER top_k 20
PARAMETER top_p 0.95
ollama create legal-27b -f Modelfile
ollama run legal-27b
The RENDERER/PARSER qwen3.5 lines matter: they enable Ollama's built-in chat template and
tool-call/thinking parsing for the Qwen3.5/3.6 family.
See the main model card for
training details, the retrieval (RAG) recommendation, and warnings (abliterated base; not legal
advice).