An open-source vertical AI-employee model from Flywheel by OpSpot, fine-tuned (LoRA) from Qwen/Qwen3.6-35B-A3B (Apache-2.0) for the legal-intake domain.

Professional law-firm intake coordinator: new-client intake, matter categorization, conflict-check fact capture, document-collection checklists, consult scheduling, fee and retainer logistics, and status updates.

Base: Qwen/Qwen3.6-35B-A3B · License: Apache-2.0 · Version: 35b-v1
Formats: safetensors (transformers / vLLM, ~65G) + model-q4_k_m.gguf (llama.cpp / Ollama, ~20G)

Download (one command)

bash

pip install -U huggingface_hub
hf download flywheel-ai/legal-intake                      # full repo (safetensors + GGUF)
hf download flywheel-ai/legal-intake model-q4_k_m.gguf    # just the GGUF

Run

bash

# llama.cpp
llama-server -m model-q4_k_m.gguf -ngl 999
# Ollama (pulls the GGUF straight from HF)
ollama run hf.co/flywheel-ai/legal-intake
# vLLM (serves the safetensors)
vllm serve flywheel-ai/legal-intake

Guardrail

Not a lawyer and never gives legal advice; legal questions are captured and routed to a licensed attorney at the firm.

Provenance & honesty

v1.0 is trained on synthetic seed data authored by permissively-licensed local models (Apache/MIT teachers only — never distilled from closed models). On general prompts it is roughly on par with the base; the niche edge sharpens as consented real usage flows through the OpSpot flywheel. Built on Qwen3.6 (Apache-2.0).