Third-generation medical LLM from Benyou Wang's FreedomIntelligence group (CUHK-Shenzhen): RL-only domain adaptation directly from base models via off-policy seeding ("SeedRL"), skipping the SFT stage that HuatuoGPT-o1 still used. Released as 8B and 32B checkpoints plus a 7B variant on a Pangu base (HuatuoGPT-3-7B-Pangu).

HuggingFace-first release — the technical report was still "coming soon" as of mid-2026; details beyond the model cards remain undisclosed, and the entry joins the paper-backfill work-list. Derivatives of open bases, not from-scratch pretrains.

Model Details

Variants

Name Parameters Notes
HuatuoGPT-3-8B RL-only adaptation of an open 8B base
HuatuoGPT-3-32B RL-only adaptation of an open 32B base
HuatuoGPT-3-7B-Pangu variant on an openPangu base
medicalreasoningpost-trainingopen-weight

Related