HuatuoGPT-3 skips fine-tuning and lands 71.4 on HealthBench
A medical model that never went through a supervised fine-tuning stage is now the strongest entry in its own family, and the lab behind it shipped the grader as well. Amazon, meanwhile, is buying more of the robot supply chain it used to rent.
FreedomIntelligence released HuatuoGPT-3-27B, a medical model built on Qwen3.8-27B that the team reports at 71.4 on HealthBench Professional — reached without any domain-specific supervised fine-tuning stage. Domain adaptation in medicine normally runs in two steps: fine-tune the base model on clinician-written answers, then align it with reinforcement learning on top. OnePO collapses that into a single RL pass. Teacher responses provide temporary guidance at the start of training and are retired as the student's own answers begin to match or exceed the teacher's reward, while an adaptive objective lets the model learn from low-probability teacher tokens it would otherwise discard.
The number move is the least interesting part. Last week the same lab shipped a 9B version on Qwen3.5-9B at 68.3 on the same benchmark, so the recipe travels across backbones. What makes this release worth your attention is that the reward machinery is open too: the training code, a 20,338-task medical RL set split between 10,191 multiple-choice questions with exact-match checking and 10,147 open-ended prompts graded against clinical rubrics, and an 8B rubric grader that evaluates several criteria in a single generation to keep training-time grading cheap. The weights are Apache-2.0 and the model is designed to run in thinking mode, producing its reasoning before the final answer. Treat 71.4 as the lab's own figure, measured with OpenAI's official HealthBench grading protocol rather than an independent run.
Our read: the supervised fine-tuning stage is the part of specialist-model building that needs doctors in the loop, writing and curating answers. If a single RL stage with a rubric and an open grader gets you a competitive medical model, the expensive input shifts from clinician hours to eval design — which is the harder thing to fake and the easier thing to compare. It also relocates the risk. A rubric grader is now a load-bearing component of a clinical model, and it is the same size as models people deploy casually. We have argued before that "just fine-tune it" is the most popular wrong answer in AI product work — How to — decide if your model needs fine-tuning — and a paper that removes the fine-tuning stage entirely is a useful stress test of that instinct.
Amazon will spend more than $100 million on a 585,000-square-foot plant in Greenwood, Indiana that makes equipment for its warehouse robots, creating 300 jobs at an average of nearly $100,000 a year. The facility will handle fabrication, robotic welding, automated powder coating and final assembly under one roof, using AWS and AI-driven manufacturing systems, and is expected to open by 2028. It is Amazon's fourth robot manufacturing site: two in Massachusetts, a third announced for Austin, Texas in August, and now Indiana — doubling the number of plants where Amazon builds its own machines in under two months.
The scale of the installed base is the context. Amazon says its US lines have produced more than one million robots, that the machines now work across more than 300 facilities worldwide, and that they assist employees with 75% of the customer orders it delivers. Amazon's vice president of worldwide economic development, Holly Sullivan, told the Wall Street Journal that the company hired hundreds of thousands of employees alongside those robots. The New York Times reported last October that internal documents showed the company planning to avoid hiring more than 600,000 US workers by 2033 through automation — documents Amazon disputed as coming from "just one team."
Our read: insourcing the factories is how the automation math gets cheaper to expand and harder to reverse. Building the machines in Massachusetts and Indiana instead of buying them means the unit cost curve belongs to Amazon, and the same supply chain that the equipment queue made scarce for everyone else is now partially owned. Hiring 300 people to build them is the announcement; the reason a fourth plant exists at all is that the fifth one gets easier.
What to watch: whether an independent evaluation reproduces HuatuoGPT-3's HealthBench score, and whether open medical models start showing up in clinical pilots rather than leaderboards.
Would you trust a medical model whose biggest training-time dependency is a rubric grader built by the same lab? Tell us in the comments.
Sources: HuatuoGPT-3 training code (GitHub)
