Baichuan-M4

A patient reports a cough and a low-grade fever, uploads a complete blood count three days later, then mentions that the antibiotic they were given upset their stomach. Scored as single-turn question answering, those are three unrelated prompts. Baichuan-M4 is built on the premise that they are one episode of care, and that connecting them is a systems problem rather than a bigger-checkpoint problem.

Developed by Baichuan AI with the THUBPM Group at Tsinghua University, M4 was unveiled in May 2026 and released jointly with Tsinghua the following month. It is organized around three pillars rather than a single network: Baichuan-Harness, a runtime that hosts both reinforcement-learning training and production deployment so the action space and execution constraints match; a core reasoning model trained inside that runtime; and a clinical tool layer supplying patient memory, evidence retrieval, and multimodal perception. Training happens once, at Baichuan, and the released system is applied to new patients and new images without per-institution refitting.

That structure is what separates M4 from the medical multimodal models it is measured against. MedGemma, MedDr, HuatuoGPT-Vision, BiMediX2 and Med-R1 are single checkpoints that consume an image and text in one forward pass. M4 integrates vision-language perception at the tool layer instead, and its dermatology component is an explicit multi-step decision process rather than one classification. The technical report defines "clinical-grade" as readiness for physician-supervised use under predefined evaluation and safety requirements, explicitly not as autonomous diagnosis or regulatory approval as a medical device.

#Key Features

• Unified training and deployment runtime: Baichuan-Harness closes the sim-to-real gap by running reinforcement learning in the same environment that serves production, and enforces action constraints that block illegal tool calls and unauthorized data access.
• Span-level reward modeling: SPAR++ anchors reward to key clinical spans rather than scoring a whole dialogue trajectory, so the model is credited for sufficient history taking and timely risk identification, not only for reaching the right conclusion. A quality gate withholds efficiency rewards until the medical quality score clears a threshold, preventing the model from truncating a consultation to close it faster.
• Two-tier patient memory: Transient session context is kept separate from a structured long-term profile, and a medical fact enters the profile only after extraction, confidence assessment, and where needed user confirmation.
• Traceable evidence retrieval: Queries are decomposed with the PICO framework and run against a six-layer source hierarchy — primary research, systematic reviews, guidelines, practical knowledge, health education, and regulatory and real-world evidence — with relevance, timeliness, and authority as reward signals.
• Tool-layer clinical perception: Document OCR, chest X-ray report generation, and an evidence-driven dermatology agent that plans a path from examination to comparison to differential, and receives no reward for a diagnosis lacking a complete evidence chain.

#Technical Details

Baichuan has not disclosed M4's base model, parameter count, or context length; the technical report is a capabilities and evaluation document. Training combines SPAR++ span-level rewards, curriculum learning that establishes initial-visit competence before mixing in follow-up scenarios, and the SAPO policy-optimization algorithm with an R3 route-replay mechanism to suppress loss spikes. A reasoning-path compression objective cuts internal chain-of-thought tokens to roughly one-sixth, freeing context for patient memory and retrieved evidence. Reported results put HealthBench at 68.6 overall with a 15.9-point lead over GPT-5.5 on the Hard subset and a hallucination rate of 3.3%; long-context clinical memory reaches 86.9 against 65.8 for Baichuan-M3; citation precision on the 657-question Baichuan-EBM set is 90.0. Multimodal evaluation covers 297 OCR samples (structured field extraction 0.914), IU-Xray chest X-ray report generation (CIDEr 0.1892, GREEN-LLM 0.8435), and a 4,893-case Fitzpatrick 17k dermatology subset mapped to ICD-11 (30.78% top-1 exact, 60.68% top-6 category).

#Applications

M4 targets physician-supervised clinical workflows: pre-visit triage, history taking, differential screening, follow-up visits, chronic disease management, and medication feedback tracking, plus structured extraction from lab reports, prescriptions, discharge summaries, and imaging reports. The consumer-facing Baixiaoyi family-doctor agent is built on it. Weights are not released; access is by API to certified medical institutions and research units, which rules out on-premises adaptation for institutions that cannot send patient data out.

#Impact

M4's argument is that reliability in continuous care comes from orchestration, persistent memory, and traceable evidence rather than from a larger checkpoint, and its widest reported margins — long-context clinical memory and citation precision — fall exactly where that argument predicts. The caveats are substantial. Every headline number is vendor-run, and the Scan-Bench consultation benchmark, built with more than 150 front-line physicians, and Baichuan-EBM are both Baichuan's own constructions with no independent replication. The dermatology evaluation is image-only, with multi-turn visual-clinical interaction left to future work. Architecture, parameter count, training-corpus composition, and license are all undisclosed, and no M4 weights have been published, in contrast to Baichuan-M2 and Baichuan-M3, which ship openly licensed checkpoints.