Spark Medical LLM V3.5 is a vertical-domain medical AI model released by iFlytek Medical on June 9, 2026. It tackles the make-or-break question for medical AI in the real world: will doctors actually use it? Physician adoption of its generated medical records reaches 91%, documentation time drops 52%, and it took first place in both of the authoritative IDC and MedBench evaluations.

Three Core Scenarios

Medical Record Generation: Time Cut in Half

Using far-field multi-speaker speech recognition, it captures content in real time during doctor-patient conversations and automatically generates structured medical records end to end.

• Record generation plausibility rate: 80%
• Physician adoption rate: 91%
• Documentation time reduced: 52%
• Projected savings of 2.8 hours per day for outpatient doctors

An outpatient doctor seeing 100 patients a day used to spend over 4 hours handwriting records. Now, 30 seconds after the conversation ends, a properly formatted record is generated automatically, and more than nine in ten doctors can use it at a glance.

AI Imaging Reads: 75% Adoption

Through multi-sequence joint spatial modeling and image-text cross-modal reasoning, it fuses recognition and report generation into a single action. The output is not just "what's on the scan" but "what it means and what to do next."

• Physician adoption for X-ray and MR report generation: 75%
• Imaging quality control reaches expert-consultation level

This is the first time a medical LLM has crossed that line in a real clinical setting anywhere in the world.

Evidence-Based Clinical Decision Assistant Agent

It follows a technology path of "evidence alignment -- reflection checking -- expert reinforcement," so every clinical recommendation can be traced back to authoritative medical literature.

• Covers medical-record content quality control, intelligent medication review, and general-practice and cross-specialty diagnostic assistance
• Connected to a large body of authoritative clinical guidelines and expert consensus
• Compatible with Chinese clinical standards and both Western and traditional Chinese medicine systems

Technology Foundation

Fully Domestic Compute

It was the first to run efficient long-context training with DSA (Dynamic Sparse Attention) and MTP (Multi-Token Prediction) on fully domestic compute platforms, raising medical long-context inference throughput 4.5x.

Data Flywheel

• 1.6 billion person-times of de-identified medical speech data
• 1.2 billion real diagnosis-and-treatment records accumulated
• Over 2.2 million new cross-modal real medical samples added daily
• Cross-modal alignment of speech, imaging, and medical-record text

Authoritative Evaluations

On multiple key tasks, Spark Medical V3.5's overall capability significantly surpasses general-purpose LLMs such as GPT-5.5, DeepSeek-V4-Pro, and Qwen3.7-Max.

Scale of Deployment

How to Access It

Hospitals

Connect through official iFlytek Medical channels; on-premises deployment and regional health-platform integration are both supported.

Consumers

Download the "iFlytek Xiaoyi" app, which supports:

• Multimodal consultations (photo-based pill identification, photographed lab-report interpretation)
• Dynamic memory management of health records
• Precise personalized consultations

Who It's For

• Clinicians at Grade 3A hospitals and primary-care institutions
• Hospital IT departments and medical AI procurement decision-makers
• Individual users who need health management and consultation
• Regional public health management platforms

💡 Tip: A 91% adoption rate means over nine in ten AI-generated records can be used directly, but doctors should still perform the final review in practice. Medicine has essentially zero fault tolerance — AI assists, it does not replace doctors.

---

Sources: iFlytek Medical official announcements, IDC reports, MedBench evaluations