Meet AntAngelMed: A 103B-Parameter Open-Source Medical Language Model Built on a 1/32 Activation-Ratio MoE Architecture
103B parameters. 6.1B activated. AntAngelMed just topped every open-source medical AI leaderboard—and it's not even close.
Why it matters
A new open-source medical LLM with efficient MoE architecture is establishing a new performance-per-compute frontier for specialized domain models, signaling how technical moats in healthcare AI are shifting toward activation efficiency rather than raw parameter count.
The key facts
8 to know103B total parameters, 6.1B activated at inference (1/32 activation ratio)
Matches performance of 40B dense models
200+ tokens/second throughput on H20 hardware
Ranks #1 on OpenAI HealthBench among open-source models
Tops MedAIBench and MedBench leaderboards
Three-stage training: continual pre-training + SFT + GRPO RL
Built on Ling-flash-2.0 architecture
Open-source release (no paywall)
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: MedAIBase has released AntAngelMed, a 103B-parameter open-source medical language model that uses a 1/32 activation-ratio Mixture-of-Experts (MoE) architecture to activate only 6.1B parameters at inference time, matching the performance of roughly 40B dense models while exceeding 200 tokens per…