We open-sourced Chaperone-Thinking-LQ-1.0 — a 4-bit GPTQ + QLoRA fine-tuned DeepSeek-R1-32B that hits 84% on MedQA in ~20GB[N]

Hey everyone, We just open-sourced our reasoning model, Chaperone-Thinking-LQ-1.0, on Hugging Face. It's built on DeepSeek-R1-Distill-Qwen-32B but goes well beyond a simple quantization — here's what we actually did: The pipeline: 1. 4-bit GPTQ quantization — compressed the model from \~60GB down to \~20GB 2. Quantization-aware training (QAT) via GPTQ with calibration to minimize accuracy loss 3. QLoRA fine-tuning on medical and scientific corpora 4. Removed the adaptive identity layer for transparency — the model correctly attributes its architecture to DeepSeek's original work Results: |Benchmark|Chaperone-Thinking-LQ-1.0|DeepSeek-R1|OpenAI-o1-1217| |:-|:-|:-|:-| |MATH-500|91.9|97.3|96.4| |MMLU|85.9|90.8|91.8| |AIME 2024|66.7|79.8|79.2| |GPQA Diamond|56.7|71.5|75.7| |MedQA|84%|—|—| MedQA is the headline — 84% accuracy, within 4 points of GPT-4o (\~88%), in a model that fits on a single L40/L40s GPU. Speed: 36.86 tok/s throughput vs 22.84 tok/s for the base DeepSeek-R1-32B — about 1.6x faster with \~43% lower median latency. Why we did it: We needed a reasoning model that could run on-prem for enterprise healthcare clients with strict data sovereignty requirements. No API calls to OpenAI, no data leaving the building. Turns out, with the right optimization pipeline, you can get pretty close to frontier performance at a fraction of the cost. Download: [https://huggingface.co/empirischtech/DeepSeek-R1-Distill-Qwen-32B-gptq-4bit](https://huggingface.co/empirischtech/DeepSeek-R1-Distill-Qwen-32B-gptq-4bit) License is CC-BY-4.0. Happy to answer questions about the pipeline, benchmarks, or deployment.