AI Briefing20GB Medical Reasoning
·2026.04.22 05:07
Key point
A 4-bit reasoning model based on DeepSeek-R1-32B has been released, reduced to 20GB.
Details
Chaperone-Thinking-LQ-1.0, based on DeepSeek-R1-Distill-Qwen-32B, has been open-sourced on Hugging Face.
The core is 4-bit GPTQ quantization and QLoRA fine-tuning.
- Compressed model size from about 60GB to 20GB using GPTQ
- Minimized accuracy loss via QAT including calibration
- Additional QLoRA training on medical and scientific corpora
- Removed the adaptive identity layer to keep the architecture's origin transparent
The published performance is as follows.
- MATH-500 91.9
- MMLU 85.9
- AIME 2024 66.7
- GPQA Diamond 56.7
- MedQA 84%
Speed was also improved.
- Throughput: 36.86 tok/s
- base DeepSeek-R1-32B: 22.84 tok/s
- About 1.6x faster, with median latency reduced by about 43%
The author stated that this model can run on a single L40/L40s GPU and was designed for on-premise medical customers who require that data never leave their premises.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Jevper Released: A Jev Classification Wrapper for OpenAI-Compatible Models with Automatic Logprobs Detection