LingSante-3.1-air-VL tops clinical benchmarks, runs locally
Ant Group's Inclusion AI released LingSante-3.1-air-VL, an open-source 124B multimodal mixture-of-experts medical model featuring 5.5B active parameters per token and a 256K context window. The model outperforms proprietary frontier models on chest X-ray and clinical reporting benchmarks, while a community W4A16 quantization allows local execution on four consumer RTX 3090 GPUs.
Open-weights clinical AI has crossed a critical threshold where high-accuracy medical reasoning no longer requires closed proprietary APIs or massive datacenter clusters. LingSante demonstrates that specialized MoE architectures can deliver frontier multimodal diagnostic performance on accessible hardware.
- –Extreme parameter efficiency: With 124B total parameters but only 5.5B active per token across 512 routed experts, the model delivers low-latency inference on modest compute budgets.
- –Hybrid attention architecture: Pairing 35 KDA linear attention layers with 7 Gated MLA layers keeps the KV cache compact, allowing full 256K context processing with minimal VRAM overhead.
- –Consumer hardware deployment: The community W4A16 quantization compresses routed experts down to 71 GB, allowing 4× RTX 3090 rigs to achieve 100-140 tokens/second without compromising vision or reasoning precision.
- –Agentic workflow integration: Built-in support for multi-turn symptom collection, medical search, and GUI tool use allows the model to retrieve evidence and execute clinical administrative workflows.
- –Domain-specific open parity: Outperforming models like GPT-6.1 Sol on MIMIC-CXR and AFUMED benchmarks offers healthcare institutions a self-hosted alternative where data privacy is paramount.
DISCOVERED
38m ago
2026-11-22
PUBLISHED
1h ago
2026-11-22
RELEVANCE
AUTHOR
0x0SojalSec