Baichuan Intelligent CEO Wang Xiaochuan used a definitive statement to outline a clear boundary for the company's direction, which had been frequently pulled in different directions by external opinions over the past two years: "We lack the capability to dabble in finance, then entertainment, then healthcare all at once; we can only deeply cultivate one main path."

Simultaneously, Baichuan released and open-sourced its new-generation medical large language model, Baichuan-M3. On the HealthBench medical AI evaluation platform introduced by OpenAI, Baichuan-M3 ranked first with a score of 65.1 points; in a pure model setting that does not rely on tools or retrieval-augmented generation, its medical hallucination rate dropped to 3.5%, achieving the current world's lowest level.

Wang Xiaochuan indicated that the company has approximately 3 billion yuan in cash on its books, signifying that Baichuan can sustain continuous investment in its chosen track. He stated that from the very day Baichuan Intelligent was founded, he had written in an all-staff letter: aspiring to contribute to the development of life sciences and medicine over the next twenty years, and to make a contribution to public health, with the core path being the construction of mathematical models for life and health, and that they have already taken action.

Discussing the recently high-profile topic of AI large model companies going public, Wang Xiaochuan commented that "they are primarily riding on the technological dividend of general models and policy support," while medical AI will mature a bit later, with a timeline of one or two more years. "Baichuan is expected to initiate its IPO listing in 2027."

Ju Qiang, the head of model technology at Baichuan Intelligent, revealed that the company currently dedicates about 80% of its computing power to reinforcement learning-related training, and the M3 model is a phased achievement formed under this training strategy. Besides strong reasoning capabilities and low hallucination rates, end-to-end consultation ability is another standout feature of M3.

Beyond the model, Baichuan also provided a timeline for its product side – in the first half of this year, Baichuan will successively release two toC medical products. Initially offered for free, paid capabilities can be introduced later by module, focusing on serving scenarios like assisted decision-making for patients and home health monitoring. "This is different from the many general health and medical AI products everyone has been seeing on the market lately," Wang Xiaochuan said.

With 80% of computing power invested in reinforcement learning, the M3 model represents a fundamental shift in training focus compared to the Baichuan-M2 released last May. Its core concept can be summarized in one phrase: fact-aware reinforcement learning.

Ju Qiang stated that medical large models universally face a difficult-to-avoid problem: models with stronger reasoning capabilities are more prone to generate hallucinations in medical scenarios; while一味 suppressing hallucinations can make the model overly conservative when faced with complex problems. During the Baichuan-M2 phase, Baichuan relied more on toolchains and post-processing methods as a safety net. Baichuan-M3, however, chooses to address the hallucination problem earlier, moving it to the training stage.

Ju Qiang further explained: M3's training does not simply increase the proportion of reinforcement learning, but rather redefines "what constitutes an error." When the model provides seemingly reasonable but factually unsubstantiated medical judgments, such outputs are explicitly penalized during training; meanwhile, the model's exploration space within the reasoning chain is not compressed. This perception of factual consistency is key to fact-aware reinforcement learning.

Centered on this goal, Baichuan-M3 implemented several key adjustments at the training and algorithm levels. The first change occurred in the dynamics of reinforcement learning. In the Baichuan-M2 stage, the patient's state was dynamic, but the "doctor evaluation model" responsible for scoring was relatively fixed; with Baichuan-M3, the evaluation model itself also iterates as the main model's capabilities improve, preventing the model from hitting a capability ceiling prematurely during later training stages.

The second upgrade is reflected in the shift in hallucination control methods. Baichuan-M3 no longer relies on external evidence-based tools to correct outputs but instead handles hallucination suppression internally within the model. This allows the consultation process to maintain continuity without frequent interruptions from tool calls.

The third change addresses the long dialogue structure unique to medical scenarios. Ju Qiang mentioned that existing general reinforcement learning algorithms tend to be unstable in multi-turn consultations; Baichuan specifically modified the algorithm structure to enable the model to maintain goal consistency over longer dialogues.

These technical adjustments endow Baichuan-M3 with "native, end-to-end serious consultation capabilities." Wang Xiaochuan emphasized that this is completely different from using prompts to make a large model play the role of a doctor.

After discussing the technology, Wang Xiaochuan spent considerable time explaining "why healthcare must be completely reinvented." In his view, there are four long-term structural deficiencies in current domestic healthcare in China. First, the number of doctors consistently fails to match demand; second, the doctor-patient relationship is highly unequal, with patients having the least information yet bearing the consequences of decisions; third, China lacks a family doctor system, causing a large number of patients to passively flock to top-tier hospitals; finally, medical science itself still has cognitive blind spots, and doctors also face situations of uncertainty and lack of knowledge.

Based on this, Baichuan wants to enable ordinary people to understand their medical situations clearly. To know what they are experiencing, why a certain approach is taken, and what choices are available for the next step. You can understand what the doctor is evaluating, you can reiterate this information to another doctor, and you know roughly what might happen if you choose option A or B.

Consequently, Baichuan chooses to focus more energy on out-of-hospital diagnosis and treatment scenarios, especially the assisted decision-making capability for patients facing uncertain symptoms at home. In Wang Xiaochuan's narrative, an important reason for Baichuan's approach is the team's belief that the real future增量 in healthcare fundamentally won't originate from within hospitals.

In Wang Xiaochuan's view, hospitals primarily承担 executive functions. "You have already decided what needs to be done; you come to the hospital for tests, surgery, medication, or monitoring. The judgments that truly influence the patient's path often occur much earlier. Should you pay attention when symptoms appear? Where should you go first? Is a second opinion needed? These decisions are frequently made without a doctor present."

This choice also directly determines its commercialization direction – it is "serious medicine" focused on "out-of-hospital needs." Although the products have not yet been unveiled, according to their responses, the launched products will not cross regulatory boundaries to provide diagnoses or prescriptions; their main functions will be to help users understand information, organize symptoms, and clarify the next steps.

Wang Xiaochuan stated that, in terms of capability, Baichuan-M3 is already sufficient to play such a role, but this does not mean Baichuan will rush to push the model into all scenarios. Baichuan's medical AI product concept covers all disease types but has defined clear priorities: the first step focuses on pediatrics and oncology. Currently, they are collaborating with Beijing Children's Hospital and the Cancer Hospital of the Chinese Academy of Medical Sciences to advance validation in real-world scenarios.