"Rallying the troops and re-establishing connections with everyone." Baichuan Intelligent CEO Wang Xiaochuan used this phrase to explain why he chose to speak out again at this juncture, stating, "Previously, when we shifted from general (models), we were heavily criticized."

Currently, the concept of the "Six Tigers of Large Models" is rarely mentioned. A parallel reality is that these six companies have indeed embarked on increasingly divergent development paths over the past two years. Baichuan Intelligent, which was the first to pivot towards vertical healthcare application scenarios, has borne the brunt of interpretations such as "falling behind" or "lagging."

Wang Xiaochuan does not shy away from the current industry landscape he faces. Commenting on the successive listings of large model companies Zhipu AI and Minimax, Wang assessed that they "capitalized on the technological红利 of general models and national support for a tech-powered nation," but believes their market valuations do not yet match their commercial capabilities.

Regarding Baichuan Intelligent's own financing plans, Wang Xiaochuan stated that the company might initiate an IPO in 2027 and currently has 3 billion yuan in cash reserves.

"Now we are prepared and can start delivering results to everyone," he said.

On January 13, Baichuan Intelligent announced the open-sourcing of its new-generation medical large model, Baichuan-M3. According to disclosed evaluation results, the model ranked first in the HealthBench medical AI benchmark test released by OpenAI; it also took the top spot in the HealthBench Hard subset, which places greater emphasis on complex clinical decision-making capabilities.

M3 is already live on "Baixiaoying," providing AI consultation services. Wang Xiaochuan stated that the company will adhere to a To C product strategy, planning to launch a new, independent C-end product in 2026 focused on serious medical scenarios.

However, under current regulatory constraints, Baichuan Intelligent will not cross the red line of providing diagnostic conclusions in the short term. The product will not directly issue diagnoses or prescriptions but will assist patients in understanding treatment options and making choices, thereby establishing its value in aiding decision-making. Wang Xiaochuan mentioned that the product will be free initially, with commercialization planned later through service packages or partnerships with pharmaceutical and medical device companies.

When asked to define a user base size that would validate the product's value, Wang intuitively suggested the number "100,000 users."

Baichuan Intelligent also disclosed the model's performance in controlling medical hallucinations. Without relying on external retrieval or tool systems, M3's medical hallucination rate is 3.5%, lower than publicly available results for mainstream general models in similar evaluations.

Ju Qiang, Baichuan Intelligent's Model Technology Lead, explained in an interview that their approach involves incorporating medical fact consistency as a core objective during the model's reinforcement learning training phase. This Fact-Aware RL architecture allows for reducing hallucinations while simultaneously enhancing reasoning capabilities.

Another key advancement of M3 is its end-to-end capability for serious medical consultations. Differing from common practices reliant on role-playing prompts, Baichuan employed native training methods centered around its proposed SCAN principle. This enables the model to proactively ask follow-up questions and prioritize risk-related information gathering, mimicking a clinician's approach.

Regarding multimodality, Wang Xiaochuan emphasized that language remains the central axis of intelligence, with image recognition serving merely as an auxiliary tool. Future plans involve integrating multimodal perception models, but the main battlefield will remain reasoning systems based on symbolic logic.

Over the past year, healthcare has become a key focus area for increased investment by general large model developers. OpenAI has launched ChatGPT Health for medical scenarios, and Anthropic released Claude for Healthcare targeting medical institutions. Against this backdrop, Baichuan Intelligent's M3 provides a new comparative benchmark in the competitive landscape of medical large models.

In this field, Wang Xiaochuan believes Baichuan Intelligent's competitive advantages stem from three aspects: consistently leading model capabilities, strategic determination to enter high-value, non-consensus scenarios, and unique product form innovation.

Addressing competition from major tech firms, specifically the recently high-profile Ant Group's A Fu, Wang categorized it as a泛健康 product. He positioned Baichuan as more targeted towards serious medicine, with the core goal of addressing clinical needs rather than providing superficial consumer-grade services.

When discussing focus areas within healthcare, Wang mentioned pediatric chronic diseases and oncology as current priorities. For long-term goals, the company aims to achieve higher-level model intelligence through training paradigm iterations and mining more data, potentially tackling unsolved complex diseases in medicine.

"Actually, I wanted to focus on healthcare from day one, but internally, perhaps influenced by media or other pressures, many colleagues were unhappy, feeling it wasn't their life's ideal," Wang Xiaochuan briefly reflected on the turmoil the company experienced recently towards the end of the communication session.

"So now I've learned my lesson. I interview every new hire personally and tell them straight: I intend to focus on healthcare," Wang Xiaochuan said.