A generated title card reading 'LingSante-3.1-flash vs Ling 3.1 Flash' with the subtitle 'Same 560B weights. Different training. What changed?' Three cards read 'HealthBench Pro 65.35 to 65.72', 'HLE 37.44 to 44.07' and 'Terminal-Bench 46.03 to 26.5, harness changed', with a footer 'Both cards vendor-reported, October and November 2026.' The OrcaRouter logo is composited along the bottom edge.
Engineering & Research

LingSante-3.1-flash vs Ling 3.1 Flash: What Medical Training Bought

Auteur

jinhao.song

Date de publication

Derniers modèles · 20Voir tous les modèles →
Benchmarks : Artificial Analysis · mis à jour quotidiennement
Retour à tous les articles

The cleanest experiment in this release is the one Ant Group ran on itself. LingSante-3.1-flash and Ling-3.1-flash share the same architecture, the same 560B total parameters, the same 25B active per token, the same tokenizer and the same MIT licence. The only difference is training: medical data mixed through pretraining, mid-training and supervised fine-tuning, then medical reinforcement learning and multi-teacher distillation. That makes the delta between the two cards the closest thing to a controlled measurement of what domain training buys.

The headline delta is 0.37 points. Ling-3.1-flash scores 65.35 on HealthBench Professional; LingSante-3.1-flash reports 65.72. After a training programme that the launch post describes as three dedicated frameworks, the medical rubric moved by less than half a point.

That is not the whole result, and it is not even the most interesting one. But it is the right place to start, because it reframes every other claim in the release.

A 0.37-point headline improvement

HealthBench Professional is the benchmark Ant leads with, and it is the one where the medical training should show up most clearly. Both numbers come from the same organisation and, per the footnotes, the same evaluation environment. The base model's card reports 65.35. The medical model's card reports 65.72.

Two readings are possible. The cynical one is that the medical training did almost nothing and the release is a rebrand. The charitable one is that the base model was already strong on this rubric — it had healthcare data in its general training mix — and a specialist can only add so much on a benchmark that measures conversational clinical quality. The second reading is better supported by the rest of the table, but neither reading makes 0.37 a compelling reason to deploy a second 560B model.

What the 0.37 does do is undercut the launch framing. The post says the model "unifies diverse medical capabilities while retaining general capabilities." On the medical rubric, unification bought a rounding error; on general capability, the next sections show what it cost.

The reasoning gain is real

The strongest evidence for the medical training is not the medical rubric. It is HLE, a general high-level reasoning benchmark, where the specialist gains 6.63 points.

• HLE, text only, no tools — Ling-3.1-flash 37.44, LingSante-3.1-flash 44.07.

• Tau3-Banking — 47.80 against 53.61, a 5.81-point gain on a tool-use benchmark that has nothing to do with medicine.

• MCP-Atlas — 83.00 against 83.98, a small gain on multi-tool orchestration.

• Automation Public v1.0.6 — 52.50 against 56.50.

• AA-LCR, long-context reasoning — LingSante reports 84.3; the base card does not publish an AA-LCR row, so there is no delta to compute.

A medical reinforcement-learning stage that lifts general reasoning by six points is doing something real. The plausible mechanism is that clinical reasoning tasks reward careful inference over long contexts, and that skill transfers. Ant's own description — reasoning, tool use, search and professional behaviour trained with contribution-aware rewards — is consistent with a model that learned to think more carefully, not merely to talk about medicine.

The transfer also shows up on BrowseComp with context, where LingSante reports 94.10 against the base card's 91.67. That comparison carries a caveat: the base card's BrowseComp row does not specify the context setting, and the LingSante row explicitly says "with context." A 2.4-point gap across different settings is directional at best.

Screenshot of the Hugging Face model card for inclusionAI/LingSante-3.1-flash captured 23 November 2026 in English: 27 likes and a blank downloads-last-month figure; tags Text Generation, Safetensors, bailing_hybrid, conversational, custom_code and fp8 with License mit; sidebar showing 572B parameters in BF16, F32 and F8_E4M3, the line 'This model isn't deployed by any Inference Provider', and a six-item LingSante collection; introduction stating 560 billion total parameters, 25 billion per token and a 1-million-token context window; the research-and-evaluation-only notice; and a grid of benchmark chart thumbnails whose legend names LingSante-3.1-flash, DeepSeek-V4.1-Flash-1017, GLM-5.3, Kimi-K3.1-Pro, Claude Opus 5.5, GPT-6.2 Astra and Gemini 4 Argon.

The agentic collapse is partly the harness

The alarming rows are the coding and terminal ones, and they are the ones most likely to be misread.

• Terminal-Bench 4.0 — Ling-3.1-flash 46.03, LingSante-3.1-flash 26.5. A 19.5-point drop.

• SWE-Bench Pro — 68.74 against 65.25, a 3.5-point drop.

• DeepSWE — 59.70 on the base card; not reported on the medical card.

• GDPval-AA v2.1 — 1,673 Elo against 1,600, a 73-point drop.

Read alone, those numbers say the medical training destroyed the model's ability to work in a terminal. Read with the footnotes, they say something narrower.

The base-model card evaluated DeepSWE, Terminal-Bench 4.0, SWE-Bench Pro and Codebase QnA using LingCode, a harness Ant built on the open-source agent Pi and has not open-sourced. The medical card evaluated SWE-Bench Pro with Claude Code v2.1.205 and Terminal-Bench 4.0 with Mini-SWE-Agent v2.4.6. Those are different harnesses measuring the same benchmarks. A 19.5-point swing between harnesses on an agentic benchmark is entirely plausible — harness choice routinely moves terminal scores by double digits — so the drop cannot be attributed to the medical training with any confidence.

The GDPval drop is harder to explain away. Both cards report GDPval-AA v2.1, an Elo-scale benchmark of economic-value tasks, and 73 points is a meaningful regression. That is the row that most credibly suggests a capability trade: the medical model is worse at general professional work than its base.

The honest verdict on the agentic rows is mixed. Some of the apparent collapse is measurement. Some of it, particularly GDPval, is probably real. Anyone quoting "medical training halved Terminal-Bench" is quoting a harness change.

Serving the same 560B twice

Because the architectures are identical, the deployment economics are identical too — which makes the comparison unusually clean and unusually unforgiving.

• Weights — both are FP8 checkpoints around 574 to 578 GB across 54 shards, with the same SGLang day-zero image and vLLM recipe. The medical model adds a 291.72 GiB MXFP4 sibling; the base model's quantised builds are community work.

• Context — both advertise one million tokens, both ship a configuration that documents 262,144, and both require operator configuration to reach the larger window.

• Hosting — Ling-3.1-flash has a list price of $0.30 input and $0.90 output per million tokens through first-party and third-party providers. LingSante-3.1-flash has no first-party list price; it is self-hosted or bought through the vendor's launch partner.

• Independent score — Ling-3.1-flash's 1019 checkpoint is at 29.6 on the Artificial Analysis Intelligence Index v5.0.1. LingSante-3.1-flash has no Index and one independent medical result, 23.3% on MLCR-AA.

So the choice is not "cheap base versus expensive specialist" in the usual sense. It is the same hardware bill twice, with the base model additionally available as a metered API. The specialist has to justify itself on capability alone, because it does not win on access or on price.

For teams weighing the general alternative rather than the base model, the comparison set is wider. DeepSeek V4.1 Flash, GLM 5.3 Flash and Gemini 3.8 Flash are all on OrcaRouter at provider list price with no markup, which makes a general-purpose baseline cheap to run beside a self-hosted specialist. Neither Ling checkpoint is on the platform, so the base-versus-specialist test has to be run on your own infrastructure.

A generated single-column delta scorecard titled 'LingSante-3.1-flash vs Ling 3.1 Flash — what changed'. Six rows read 'HealthBench Pro: plus 0.37', 'HLE no tools: plus 6.63', 'Tau3-Banking: plus 5.81', 'GDPval-AA: minus 73 Elo', 'Terminal-Bench 4.0: minus 19.5, harness changed' and 'Licence and size: MIT, 560B total, 25B active, both'. A footer reads 'Both cards vendor-reported; coding rows used different harnesses.'

Which checkpoint belongs in production

If the workload is clinical judgement on text, LingSante-3.1-flash is the better checkpoint. The 44.07 HLE score, the 53.61 Tau3-Banking score and the medical-agent rows — EHR-Complex 74.40, PatientAgentBench 4.722, PhysicianBench 57 — describe a model that reasons more carefully about complex cases. The 0.37-point HealthBench delta undersells that; the reasoning benchmark overstates it less.

If the workload is general agentic work — terminal tasks, code review, long-running tool chains — Ling-3.1-flash is the safer choice. Its base-card numbers on Terminal-Bench 4.0 and SWE-Bench Pro are higher, its GDPval score is higher, and its harness is the one Ant designed for its own coding agents. Some of that advantage is measurement, but not all of it.

The case that does not hold up is running both. They share a 578 GB footprint and a serving stack, so running them side by side doubles the hardware for a capability delta that is real but narrow. A single deployment with routing at the prompt level — clinical cases to the specialist, general agent traffic to the base — only makes sense if the two models are behind one gateway, which is an engineering project rather than a configuration flag.

What the release actually demonstrates is narrower than the announcement implies and more useful than the 0.37 suggests: medical reinforcement learning on a strong base model improves careful reasoning across domains, modestly improves medical conversation, and probably costs some general professional-task performance. That is a normal trade for a domain model. The mistake would be to read it as a free upgrade.

Screenshot of the Hugging Face model card for inclusionAI/Ling-3.1-flash captured 23 November 2026 in English: 178 likes and 5,936 downloads last month; 572B parameters in BF16, F32 and F8_E4M3; an inference-provider panel in the sidebar; introduction stating 560B total and 25B active parameters with a 1M-token context window; a grid of benchmark thumbnail charts whose legend names Ling-3.1-flash, GPT-5.6 Sol, Claude Opus 5, Kimi K3, GLM-5.3, GLM-5.3 Flash and DeepSeek-V4.1-Flash; and an evaluation sidebar showing DeepSWE 59.7, SWE-bench Pro 68.74 and HLE 37.44.