A generated title card reading 'LingSante-3.1-flash' with the subtitle 'A 560B medical model, uploaded without an announcement'. Three rounded cards below read '19 Nov 2026 - repository created', '0 downloads, 1 like' and 'MIT - 54 weight shards'. A wide panel labelled 'What the vendor published' carries the line '20 medical benchmark rows, no independent run', and a footer reads 'Status as of 22 November 2026 - weights live, announcement absent.' The OrcaRouter logo is composited in a white strip along the bottom edge.
Engineering & Research

LingSante-3.1-flash: A 560B Medical MoE Nobody Announced

Autore

jinhao.song

Data di pubblicazione

Ultimi modelli · 20Vedi tutti i modelli →
Benchmark: Artificial Analysis · aggiornato ogni giorno
Torna a tutti gli articoli

At 04:30 UTC on 19 November 2026, the inclusionAI organisation on Hugging Face created a repository called LingSante-3.1-flash. It holds fifty-four safetensors shards, roughly 578 GB of FP8 weights, an MIT licence, and a model card describing a 560-billion-parameter mixture-of-experts system built for medical reasoning and clinical workflows. There was no blog post. The Ant Ling developer site, where the team normally publishes launch notes, has no entry for it as of 22 November. The repository has recorded zero downloads and one like. The predecessor medical model, Ling-3.0-flash-Sante, has 2,769 downloads and 87 likes.

That gap between a complete weight release and an empty announcement channel is the story. LingSante-3.1-flash exists, can be downloaded, and can be served today if you have the hardware. What does not exist is a second opinion about it. Every performance number in this article comes from the vendor's own card, and the card is the only place those numbers appear. This piece separates what the repository proves from what the vendor asserts, because with no launch note and no independent evaluation, that distinction is most of what a reader can act on.

Cosa contiene effettivamente il repository

The file listing is more informative than the prose card, so start there.

• Parameters — 560 billion total, 25 billion active per token. The sidebar reports 572B, which counts the multi-token-prediction head; the card's own figure is 560B.

• Architecture — 72 layers: 63 Ki​mi Delta Attention linear layers and 9 gated MLA layers in a 7:1 ratio, plus four dense layers. 512 routed experts and one shared expert, eight routed experts activated per token. Hidden size 4,096, 64 attention heads, vocabulary 157,184.

• Lineage — the card states it is built on Ling-3.1-flash, the 560B general model that inclusionAI open-weighted on 15 October 2026. The architecture is unchanged. The difference is training: medical and general data mixed through pretraining and mid-training, then supervised fine-tuning, medical reinforcement learning, and multi-teacher on-policy distillation.

• Files — weights, config, tokenizer files, a chat template, two custom modelling Python files, and the README. No training-data manifest, no evaluation harness, no technical report, no separate paper.

• Licence — MIT, the same as the base model, with no acceptable-use rider. The card adds a blunt restriction of its own: research and evaluation only, not a medical device, not a basis for diagnosis or prescribing.

• Runtimes — a day-zero SGLang image (lmsysorg/sglang:dev-ling-3.1-flash) and a vLLM recipe that requires --trust-remote-code because the modelling code ships in the repository.

• Hardware — the quickstart launches with tensor parallelism 8 and expert parallelism 8 on eight H200s, or TP4/EP4 on four GB300s.

Screenshot of the Hugging Face model card for inclusionAI/LingSante-3.1-flash captured 22 November 2026 in English: header with one like and tags Text Generation, Safetensors, bailing_hybrid, conversational, custom_code, fp8 and License mit; sidebar showing 572B params in BF16/F32/F8 E4M3 and the line 'This model isn't deployed by any Inference Provider'; introduction stating 560 billion total parameters, 25 billion per token and a 1-million-token context window; a research-and-evaluation-only notice; a benchmark chart thumbnail; and the claim that it ranks second on medical professional behaviour behind only GPT-6 Astra.

La tabella dei benchmark, leggila attentamente.

The evaluation section is a single embedded image, 2,760 by 3,308 pixels. It is a genuinely large table: twenty benchmark rows across medical reasoning, healthcare agents, safety and ethics, search agents, professional knowledge work, tool use, and coding agents. The comparison columns are DeepSeek-V4.1-Flash-1017, GLM-5.3, Kimi-K3.1-Pro, Tencent Hy4, MiniMax-M3.1, Claude Opus 5.5, GPT-6 Astra, and Gemini 4 Argon, all at their highest reasoning setting where one exists.

Selected rows, vendor-reported and unreproduced:

• MedQA-USMLE — LingSante-3.1-flash 94.50, against Claude Opus 5.5 at 97.25, GPT-6 Astra at 97.33, and Gemini 4 Argon at 97.33.

• HealthBench Professional — 65.72, the highest figure in that row. GPT-6 Astra is 64.18, Claude Opus 5.5 is 61.19, and Kimi-K3.1-Pro is 55.87.

• MedXpertQA-Text — 57.27, behind Claude Opus 5.5 at 67.14 and Gemini 4 Argon at 71.35.

• PatientAgentBench — 4.722 on the card's five-point scale, ahead of GPT-6 Astra at 4.438 and Gemini 4 Argon at 4.368.

• PhysicianBench — 57, one point behind Gemini 4 Argon at 58.

• BrowseComp with context — 94.10, the highest in the row, ahead of GPT-6 Astra at 91.50 and Claude Opus 5.5 at 90.80.

• Terminal-Bench 4.0 — 26.5, against Kimi-K3.1-Pro at 67.2 and Claude Opus 5.5 at 59.6.

• SWE-Bench Pro — 65.25, against Claude Opus 5.5 at 89.90.

• GDPval-AA v2.1 — 1,600 Elo, against Claude Opus 5.5 at 1,866 and Kimi-K3.1-Pro at 1,861.

The card's summary claims the model ranks "second overall behind only GPT-6 Astra" on medical professional behaviour. That claim has no defined composite behind it on the card. The rows closest to it are PhysicianBench, where LingSante-3.1-flash is second to Gemini 4 Argon by one point, and PatientAgentBench, where it is first. A reader should treat the sentence as a vendor framing rather than a measurement until the card names the aggregate.

Where the 1M context claim gets fuzzy

The introduction promises a one-million-token context window, and the specification table repeats it. The configuration file that ships with the weights sets max_position_embeddings to 262,144. The MXFP4 sibling released three days later is more explicit still: it describes a "configured 262,144-token context window" and leaves the million-token figure out of its own specification table.

The evaluation notes do say a one-million-token window was adopted for BrowseComp, HLE with tools, DeepSeekQA, SWE-Bench Pro and Terminal-Bench. So the million-token mode clearly exists in the vendor's harness. What the repository does not contain is the serving configuration that enables it — no YaRN scaling block, no rope-scaling argument in the quickstart, no note explaining the gap. The practical reading is that 262,144 tokens is what you get from the shipped config and one million requires a serving recipe the vendor has not published. Plan for 262K and treat 1M as a benchmark-harness capability.

The MXFP4 release changes the hardware bill

At 03:19 UTC on 22 November, inclusionAI created a second repository: LingSante-3.1-flash-MXFP4. It is the more operationally interesting artefact of the two.

• Size — 291.72 GiB for the whole checkpoint, against roughly 578 GB for the FP8 release. That is the difference between a four-node deployment and an eight-node one.

• Method — round-to-nearest quantization with no calibration and no additional training. MXFP4 weights apply to the routed experts' gate, up and down projections; the dense MLP, shared expert, KDA and attention matrices stay in FP8; embeddings, the language-model head, router weights, norms and multi-token-prediction layers stay in BF16; a few small tensors stay in FP32.

• Recommended hardware — four H20-3e GPUs at TP4/EP4. Expert activation can run at W4A8 or W4A16 at serving time.

• Caveat in the card itself — the benchmark figure reproduced on that page belongs to the source model, not to the quantized checkpoint. Nobody has published what RTN costs on these twenty rows.

An MXFP4 release landing quietly three days after the FP8 weights is the strongest signal that this is a real deployment target rather than an abandoned upload. Quantization work is not something a team does for a model it does not intend to serve.

Screenshot of the Hugging Face model card for inclusionAI/LingSante-3.1-flash-MXFP4 captured 22 November 2026 in English: header tags 8-bit precision and fp8 with License mit; a Safetensors summary listing F32, BF16, F8_E4M3 and I8 tensor types; the line 'This model isn't deployed by any Inference Provider'; introduction stating approximately 560 billion parameters and a configured 262,144-token context window; a note that this version uses Round-to-Nearest post-training quantization with MXFP4 weights for routed experts; and a specification table beginning with source model and a checkpoint size of 291.72 GiB.

What is confirmed and what is not

Confirmed by inspection: the weights exist and are complete; the licence is MIT; the architecture matches the base model; the custom modelling code is present; the SGLang and vLLM paths are documented; the MXFP4 checkpoint exists and states its own size and quantization scheme; the config ships a 262,144-token window. Confirmed by the vendor, not by anyone else: every benchmark number above, including the twenty-row table and the "second overall" claim.

Not confirmed: that the medical training improves general capability rather than trading it away; that the million-token window works outside the vendor's harness; that RTN quantization preserves the published scores; that any of this transfers to clinical use. There is no independent evaluation, no Artificial Analysis entry, no arena result, and no third-party reproduction of a single row. The Hugging Face download counter reads zero, which is consistent with a three-day-old upload and also consistent with nobody trying.

Why a quiet ship still matters

Open-weight medical models at this scale are rare. The previous Sante release, Ling-3.0-flash-Sante, was 124 billion parameters with 5.1 billion active — a specialist built on a mid-sized backbone. LingSante-3.1-flash puts the same medical training on a 560B frontier-class base and ships it under MIT. For a hospital group, a health insurer, or a medical-software vendor that wants to keep patient data inside its own perimeter, that is a different category of artefact from a 7B medical fine-tune. It is also, at 292 GB minimum, a different category of hardware bill.

The release pattern is familiar from this lab. Ling-3.1-flash itself appeared as weights first and a developer blog later; the 1019 checkpoint followed the same way. A missing announcement is therefore weak evidence that none is coming. The things to watch are specific: a post on the Ant Ling developer blog, an entry in a neutral medical or general leaderboard, a download count that moves, a community GGUF or AWQ build, and a hosted price. Any one of those would turn this from a describable checkpoint into a testable model.

Until then, the honest description is narrow. LingSante-3.1-flash is a real, complete, MIT-licensed 560B medical mixture-of-experts release with a vendor benchmark table, a working serving path, a quantized sibling, and no outside verification whatsoever. The question is not whether the model is good. It is whether anyone outside Ant Group will run it before the vendor says a word.

A generated single-column scoreboard titled 'LingSante-3.1-flash - the scoreboard'. Six rows read 'Total parameters: 560B, 25B active', 'Context: 262,144 shipped, 1M claimed', 'Medical rows: 20, all vendor-run', 'Independent runs: none', 'Weights: 578 GB FP8, 292 GB MXFP4' and 'Licence: MIT'. A footer reads 'All figures from the vendor model card, 22 November 2026; unreproduced.' The OrcaRouter logo is composited in a white strip along the bottom edge.

Confrontati in questo articolo2

Rilevato da questo articolo · Benchmark: Artificial Analysis · aggiornato ogni giorno