Kos-1 Lite: SOTA Medical Model

Today we're announcing Kos-1 Lite, our first medical reasoning model. Kos-1 Lite achieves SOTA on HealthBench Hard at 46.6%.

Try interacting with the model here.

Motivation

General-purpose LLMs are not fit for clinical assistance. They are sycophantic when they should push back, verbose when a focused answer is needed, and lack the groundedness required to deliver sensitive medical guidance. This is expected. If you want your model to follow instructions to perform well on SWE-Bench style coding tasks, you necessarily accept some level of sycophancy. In domains like medicine, this tradeoff is unacceptable.

To fill this gap, we trained Kos-1 Lite. It is a frontier reasoning model that prioritizes clinical accuracy, reduced sycophancy, triage and deferral, and compassionate bedside manner.

Results

Kos-1 Lite achieves SOTA on HealthBench Hard and trails only GPT-5 on HealthBench. We note that these results come at a fraction of the serving costs of frontier (> 1 trillion parameter) models.

HealthBench was developed by OpenAI with 262 physicians across 60 countries to evaluate clinical reasoning beyond multiple-choice recall, using realistic open-ended conversations. The benchmark is challenging and unsaturated - frontier models like Claude Opus 4.6 and Gemini 3.1 Pro score just 20.4% and 19.8% on HealthBench Hard.

Examples

Here we compare Kos-1 Lite with GPT-5.2 on real medical queries.

Training Objective

We trained Kos-1 Lite against a reinforcement learning objective to achieve accurate clinical reasoning and a precise bedside manner.

For a group of rollouts $\{o_i\}_{i=1}^{G}$ sampled from rollout policy $\pi_{\text{roll}}$, we maximize:

$$\mathcal{J}(\theta) = \frac{1}{\sum_{i=1}^{G}|o_i|}\sum_{i=1}^{G}\sum_{t=1}^{|o_i|}\left[w_{i,t} \cdot \text{sg}\!\left(\min\!\left(\frac{\pi_\theta(o_{i,t} \mid o_{i,<t})}{\pi_{\text{old}}(o_{i,t} \mid o_{i,<t})},\;\varepsilon_{\text{high}}\right)\right) \cdot \hat{A}_i \cdot \log \pi_\theta(o_{i,t} \mid o_{i,<t})\right]$$

where:

$$w_{i,t} = \min\!\left(\frac{\pi_{\text{old}}(o_{i,t} \mid o_{i,<t})}{\pi_{\text{roll}}(o_{i,t} \mid o_{i,<t})},\; C_{\text{IS}}\right)$$

We incorporated the following:

• Decoupled three-policy objective¹ with CISPO² clipping on $\frac{\pi_\theta}{\pi_{\text{old}}}$ and token-level truncated importance sampling³ on $w_{i,t} = \frac{\pi_{\text{old}}}{\pi_{\text{roll}}}$ to correct training-inference mismatch
• No sampling adjustments⁴ with rollouts set to temp=1, top-p=1, top-k=-1
• Rollout Routing Replay (R3)⁵ to stabilize expert routing for MoE
• Mean-centered advantages without std normalization⁶, token-level loss⁷ to prevent long rollouts from being down-weighted, and zero-variance filtering⁷ to exclude groups with identical reward across rollouts

Length Control

Unlike agentic models that complete tasks over long horizons, health advice is delivered to patients and providers as text to read. We created a novel length bonus that controls thinking and answer length separately. Beyond latency and KV constraints on total context, the time to read the answer (after </think>) was a major UX problem. We intend to publish details with controlled ablations in follow-up work.

We build on DAPO's⁷ Soft Overlong Punishment — a linear penalty that decays over a window between $L_{\text{free}}$ and $L_{\text{max}}$. We extend it with three changes:

1. We assign a bonus rather than a penalty. With a penalty, a correct long answer can score lower than an incorrect short one, letting the model reward-hack into short wrong answers. A non-negative bonus preserves the reward ordering.

2. We apply the bonus independently to total tokens (thinking + output) and to output tokens only, with separate tuning for each. The model can reason at length internally while being rewarded for compressing its user-facing answer.

3. Both bonuses are zeroed when the reward score falls below a quality threshold. Without this, the model learns to write short, wrong answers.

Training details

We post-trained a medium sized MoE language model (~100B) using Miles (SGLang/Megatron).

What's Next

We are supplementing early benchmarking with physician and user review across a broader range of clinically relevant tasks and safety evaluations. We are excited to release upgraded checkpoints as training continues, and will continue to prioritize clinical accuracy, physician support and user friendliness.

Kos is the birthplace of Hippocrates.

References

¹ Hilton, J., Cobbe, K., & Schulman, J. (2021). Batch Size-Invariance for Policy Optimization. arXiv:2110.00641

² MiniMax (2025). MiniMax-M1: Scaling Lightning Attention to 128K — with 4096K Context Grounding. arXiv:2506.13585

³ Liu, J. & Li, Y. (2025). Your Efficient RL Framework Secretly Brings You Off-Policy RL Training. Link

⁴ Fang, Q. & Khazi, D. (2025). Mismatch Praxis: Rollout Settings and IS Corrections. Link

⁵ Ma, Y. et al. (2025). R3: Stabilizing MoE Reinforcement Learning by Aligning Training and Inference Routers. arXiv:2510.11370

⁶ Liu, Z. et al. (2025). Understanding R1-Zero-Like Training: A Critical Perspective. arXiv:2503.20783

⁷ Yu, Q. et al. (2025). DAPO: An Open-Source LLM Reinforcement Learning System at Scale. arXiv:2503.14476

⁸ Arora, R. K. et al. (2025). HealthBench: Evaluating Large Language Models Towards Improved Human Health. arXiv:2505.08775