This observation covers verified AI technology events formally released or materially updated in the roughly 24 hours ending at 7:48 am on 23 November 2026 in Sydney, Australia. One item is included.
InclusionAI releases the open-weight LingSante 3.1 medical model family
InclusionAI formally released the LingSante 3.1 medical model family on 22 November. The corresponding Hugging Face weight repositories first appeared on 19 November; the new fact within this observation window is the confirmed release status and the delivery of downloadable text and vision-language checkpoints. Both models use the MIT licence. Open weights do not amount to medical-device approval or clinical validation.
LingSante-3.1-flash is a hybrid-linear mixture-of-experts text model with 560 billion total parameters, 25 billion activated per token and a one-million-token context limit. Its model card provides local inference examples for SGLang and vLLM; the SGLang example uses eight H200 GPUs or four GB300 GPUs. These configurations are supplied by the developer and are not established minimum requirements for every hardware platform. The model card limits use to research and evaluation and says the output must not be used as the basis for diagnosis, prescribing or treatment.
LingSante-3.1-air-VL is a vision-language model with 124 billion total parameters and 5.5 billion activated per token. It accepts text and images, produces text and has a 256,000-token context limit. Downloadable weights are available, but there was no hosted inference service or public independent evaluation at release. Its performance on visual medical tasks therefore remains supported only by developer material, with no public replication across institutions, devices or clinical settings.
The developer also reported results across medical question answering, tool use and retrieval benchmarks. Those figures come from the releasing organisation's tests. Artificial Analysis has published one independent MLCR-AA medical long-context reasoning evaluation for LingSante-3.1-flash, recording a score of 23.3%. This test covers a single long-context medical reasoning benchmark. It does not replace clinical validation or independently confirm the developer's other reported scores. Outstanding evidence includes third-party results for the vision-language model, stability across inference frameworks and hardware, and safety and effectiveness in real medical workflows.
Sources
InclusionAI, LingSante-3.1-flash model card, repository created 19 November 2026 and release status confirmed 22 November 2026, https://huggingface.co/inclusionAI/LingSante-3.1-flash
InclusionAI, LingSante-3.1-air-VL model card, repository created 19 November 2026 and release status confirmed 22 November 2026, https://huggingface.co/inclusionAI/LingSante-3.1-air-VL
Artificial Analysis, LingSante 3.1 Flash independent analysis, accessed 22 November 2026, https://artificialanalysis.ai/models/lingsante-3-1-flash-1119