医疗大模型，最核心的十家公司！

TL;DR - A WeChat industry roundup profiling 20 companies (10 overseas, 10 Chinese) building medical large language models, arguing healthcare AI has shifted from single-task imaging/diagnosis tools to full-workflow "large-model-driven" systems. It matters as a market map of who controls models, medical data, clinical scenarios, and hospital ecosystems.

• Overseas players split by layer: foundation model vendors (Anthropic's Claude for Life Sciences / Claude for Healthcare, OpenAI's ChatGPT Health and ChatGPT for Healthcare claiming ~230M weekly health interactions, Google's Gemini/MedLM/MedGemma), ambient clinical documentation (Abridge, Ambience, Suki), evidence retrieval (OpenEvidence, licensed NEJM/JAMA/NCCN content), non-diagnostic agents (Hippocratic AI's Polaris 3.0 with a safety sandbox), and data infrastructure (Innovaccer).
• Chinese players emphasize benchmark leadership and deployment scale: Baichuan-M3 Plus claims 2.6% factual hallucination via evidence anchoring and 70% lower API cost; iFlytek's Spark Medical V3.5 reports 98.9 on MedBench and coverage of 811 counties / 77k grassroots institutions; Ping An claims top HealthBench Hard score; MedGPT reports a top CSEDB result; Ant/Alibaba claims 5,000+ public hospitals.
• Recurring commercial pattern: near-term ROI comes from documentation, coding, triage, and follow-up rather than autonomous diagnosis, with EHR integration (Epic, Cerner) as the key distribution moat.
• The piece frames competition as moving from raw model capability toward safety, regulatory compliance (FDA, EU AI Act, Chinese medical AI rules), and demonstrable clinical value; note it is a vendor-profile listicle with self-reported metrics and explicitly disclaims investment advice, and several cited dates/events are forward-looking (2026).