Skip to content
Ant's LingSante-3.1 Medical Models Released Quietly on Hugging Face Without Vendor Announcement

Ant's LingSante-3.1 Medical Models Released Quietly on Hugging Face Without Vendor Announcement

First seen 22 Nov 2026, 17:14 UTC • •

Article Content

Browse articles
ThreatCluster AI
ThreatCluster •November 22, 2026 at 19:14 UTC
  • •Ant's inclusionAI published LingSante-3.1-flash (560B MoE, MIT licence) and related medical models on Hugging Face in November 2026 without any vendor announcement.
  • •Performance claims rest solely on the vendor's model cards; no independent evaluation has been reported.
  • •LingSante-3.1-flash needs about 292 GB quantized and multiple datacenter GPUs, limiting deployment to well-resourced organisations.
  • •Repositories had near-zero downloads as of 22 November, and no security vulnerabilities or misuse are reported.
  • •All four articles come from one outlet (Orcarouter.Ai), so the account is uncorroborated by independent reporting.

On 19 November 2026, the inclusionAI organisation on Hugging Face published LingSante-3.1-flash, a 560-billion-parameter mixture-of-experts text model described in its model card as built for medical reasoning, under an MIT licence. Per the articles, a companion vision model, LingSante-3.1-air-VL, was created on Hugging Face at 04:31 UTC on 19 November, one minute after its text-only sibling. A signal encoder, LingSante-Pulse-1.0, followed on 20 November; it is a 16.7-million-parameter model that converts physiological waveforms into embeddings. Articles report that the repositories contain complete weights and configuration files, with LingSante-3.1-flash at roughly 578 GB in FP8 and LingSante-3.1-air-VL at 249.7 GB in bfloat16. As of 22 November, the Ant Ling developer site had no launch entry for these models, and no press release, benchmark write-up, or independent evaluation was found. The repositories showed very low engagement, with LingSante-3.1-flash at zero downloads and one like and LingSante-3.1-air-VL at two likes and zero downloads. All performance figures cited come from the vendor's own model cards and have not been independently verified. The articles describe the models as deployable by organisations with sufficient hardware, but they do not report any security vulnerability, exploitation, or misuse. The Orcarouter.Ai coverage is the sole source for this cluster.

Start a free Starter trial for enhanced analysis

Ask AI about this cluster

Timeline

2026-11-19
LingSante-3.1-flash repository created on Hugging Face
The inclusionAI organisation uploaded a 560B-parameter medical MoE text model with fifty-four FP8 safetensors shards and an MIT licence, with no accompanying blog post.
Orcarouter.Ai
2026-11-19
LingSante-3.1-air-VL repository created on Hugging Face
The vision-language medical model repository was created at 04:31 UTC, one minute after its text-only sibling, holding 249.7 GB of bfloat16 weights.
Orcarouter.Ai
2026-11-20
LingSante-Pulse-1.0 signal encoder uploaded
A 16.7-million-parameter encoder converting physiological waveforms into embeddings was uploaded, reportedly without the code needed to run it.
Orcarouter.Ai
2026-11-22
No vendor announcement found as of 22 November
The Ant Ling developer site had no entry for the models, and no independent benchmark or write-up was found.
Orcarouter.Ai

More articles in this cluster (4)

Following this threat?

Track Ant in your own feed — you're alerted when they show up in new reporting, leak sites or exploitation.

Free account · no card needed

Common questions

Can my organisation run these models?
LingSante-3.1-flash needs roughly 292 GB quantized and multiple datacenter GPUs, while the air-VL model is 249.7 GB in bfloat16. Deployment is practical mainly for well-resourced organisations.
Are there security vulnerabilities or active exploitation?
The articles report no vulnerabilities, exploitation, or misuse associated with these models. The concern is the absence of announcement and independent evaluation rather than a known threat.
How reliable are the performance claims?
The performance figures come only from the vendor's own model cards and have not been independently evaluated or confirmed by any outside source.