A generated title card reading 'LingSante-3.1-flash vs GLM 5.3 Flash' with the subtitle 'Only the flagship GLM was on the card.' Three cards read 'HealthBench Pro 65.72 vs 49.07', '320B total, 18B active GLM Flash' and 'GLM $0.15 / $0.50 per 1M', with a footer 'GLM rows from Ant's Ling-3.1-flash card, October 2026.' The OrcaRouter logo is composited along the bottom edge.
Guides & Insights

LingSante-3.1-flash vs GLM 5.3 Flash: Only the Big GLM Was Tested

Tác giả

Elias Hawthorne

Ngày đăng

Mô hình mới nhất · 20Xem tất cả mô hình →
Benchmark: Artificial Analysis · cập nhật hằng ngày
Quay lại tất cả bài viết

The LingSante-3.1-flash model card lists nine comparison models across 41 benchmarks. One of those columns is GLM-5.3 at maximum reasoning effort. The cheaper, faster sibling — GLM 5.3 Flash — is not on the card at all. So the matchup a cost-conscious team actually faces, a 560B open medical specialist against a 320B open multimodal workhorse at a fifth of the price, has to be assembled from two vendor tables that were never designed to meet.

That is not a fatal problem, because both tables exist and one of them overlaps. Ant's own Ling-3.1-flash card, published in October, includes a GLM-5.3-Flash column with a dozen shared rows. Combined with the LingSante card's GLM-5.3 column, it is possible to bound the Flash model's position without pretending the comparison was measured directly.

The column that is not there

Start with what the LingSante card does say about the GLM family, because it is more interesting than a missing row usually is.

• HealthBench Professional — LingSante-3.1-flash 65.72, GLM-5.3 51.18. A 14.5-point gap.

• PhysicianBench — 57 against 36. A 21-point gap, the largest between these two models on any medical-behaviour row.

• MedSCE, a safety aggregate — 83.11 against 79.94, a narrower specialist win.

• MedEthiAlign — 80.85 against 85.84. This one GLM-5.3 wins by five points, and it is the only medical-ethics row where the specialist is clearly behind.

• Terminal-Bench 4.0 — 26.5 against 41.9, a 15-point loss.

• SWE-Bench Pro — 65.25 against 60.88, a specialist win.

• GDPval-AA v2.1 — 1,600 Elo against 1,644, a GLM win.

• BrowseComp with context — 94.10 against 68.33, a 25.8-point specialist win that says more about the harness than about either model's search ability.

So against the flagship, LingSante wins the medical rubric and the search rows, loses terminal execution and economic-value work, and loses medical ethics. The shape is familiar from the other matchups in this series: judgement goes to the specialist, execution goes to the generalist.

Bounding GLM 5.3 Flash from its own card

The Flash variant is a different model, not a quantised flagship. It has 320B total parameters and 18B active, against the flagship's much larger footprint, and it is natively multimodal with text, image and video input. It launched on 26 August 2026 under an MIT licence at $0.15 input and $0.50 output per million tokens — roughly a tenth of the flagship's price.

Artificial Analysis puts GLM-5.3-Flash at 27.8 on the Intelligence Index v5.0.1, against 31.7 for GLM-5.3. That four-point gap is the best available estimate of what the Flash gives up in general capability, and it lines up with the vendor's own framing of the model as a Pareto-frontier cheap option rather than a like-for-like replacement.

The overlapping table comes from the Ling-3.1-flash card, which evaluated GLM-5.3-Flash directly:

• HealthBench Professional — GLM-5.3-Flash 49.07, in the same AQ evaluation environment used for every column on that card. Ling-3.1-flash scored 65.35 there, and LingSante-3.1-flash reports 65.72 on its own card.

• HLE without tools — 39.90, against Ling-3.1-flash 37.44 and LingSante 44.07.

• SWE-Bench Pro — 63.06, against Ling-3.1-flash 68.74 and LingSante 65.25.

• Terminal-Bench 4.0 — 32.80, against Ling-3.1-flash 46.03 and LingSante 26.5.

• GDPval-AA v2.1 — 1,641, against Ling-3.1-flash 1,673 and LingSante 1,600.

• MCP-Atlas — 81.00, against Ling-3.1-flash 83.00 and LingSante 83.98.

• Tau3-Banking — 47.20, against LingSante 53.61.

Read across the two cards and a defensible picture appears. On the medical rubric, LingSante is 16 points ahead of GLM-5.3-Flash. On general reasoning and coding-agent rows, the two are close, with GLM-5.3-Flash ahead on terminal work and behind on search and banking agents.

A generated two-column scoreboard titled 'LingSante-3.1-flash vs GLM 5.3 Flash — the scoreboard'. Left column values read 'HealthBench Pro: 65.72', 'HLE no tools: 44.07', 'Terminal-Bench 4.0: 26.5', 'MedEthiAlign: 80.85', 'Total params: 560B', 'Price: self-host only'. Right column values read 'HealthBench Pro: 49.07', 'HLE no tools: 39.90', 'Terminal-Bench 4.0: 32.80', 'MedEthiAlign: not published', 'Total params: 320B', 'Price: $0.15 / $0.50'. A footer reads 'GLM rows from Ant's Ling-3.1-flash card; LingSante rows from its own card.'

MedEthiAlign is the honest loss

The most useful row in the whole comparison is the one the specialist loses. MedEthiAlign is an internal Ant benchmark for medical ethics, and GLM-5.3 scores 85.84 against LingSante's 80.85. Tencent Hy4 scores 88.83 and Claude Opus 5.5 scores 88.08. LingSante is fifth of nine.

That matters because ethics alignment is not a peripheral capability in a clinical product. A model that reasons well about a diagnosis but handles consent, capacity or end-of-life framing less reliably is a liability in exactly the workflows where a medical specialist is supposed to be safer than a generalist. The launch post's "professional behavior alignment" claim is supported on PhysicianBench and PatientAgentBench, and undercut on MedEthiAlign. Both can be true; the aggregate "second overall" framing hides the split.

The Flash variant's ethics score is not published on either card, so there is no way to say whether the cheaper GLM inherits the flagship's advantage. Given that the Flash is 27.8 on the general Index against the flagship's 31.7, assuming it matches the flagship on a specialised ethics benchmark would be optimistic.

Two MIT MoEs, two hardware bills

Both models are MIT-licensed open weights, which makes them unusually comparable on deployment. The difference is what they cost to run.

• LingSante-3.1-flash — 560B total, 25B active, 578 GB in FP8 or 291.72 GiB in MXFP4. Eight H200s, or four H20-3e cards at MXFP4.

• GLM-5.3-Flash — 320B total, 18B active, roughly 328 GB in FP8. A materially smaller footprint, and the vendor reports about 4.4 times smaller KV cache than the flagship.

• Price if you rent tokens — GLM-5.3-Flash $0.15/$0.50 with $0.03 cached input. LingSante has no first-party list price; the vendor's launch partner gates free access behind a $20-per-month token plan.

• Speed — GLM-5.3-Flash is reported around 49 to 50 output tokens per second, about half the flagship's throughput. LingSante's card publishes no throughput figure.

The practical consequence is that GLM-5.3-Flash is the cheaper model to self-host and the only one of the two available as a metered API at list price. LingSante's advantage has to be large enough to justify a bigger cluster or a separate procurement, and on the medical rubric it is — 16 points is a lot. On general work it is not.

For teams that want to test the cheap side first, GLM 5.3 Flash is on OrcaRouter at $0.15/$0.50 with no markup, which makes it a low-cost baseline to run against a self-hosted specialist. The routing question — send medical-judgement calls to the big model and bulk extraction to the cheap one — is the same architecture the other matchups in this series arrive at.

Screenshot of the OrcaRouter model page for z-ai/glm-5.3-flash captured 23 November 2026 in English: header tags Vision, Tools, JSON and Reasoning with a 2026-08-26 date and a description stating 320B total and 18B active parameters, 1M-token context, text+image+video input and text output; a specification panel showing 1M context, 128K max output and p50 TTFT 6.94 s; price tiles reading $0.15 input and $0.50 output per 1M tokens; a traffic figure of 9,393.7M tokens over seven days; and a pricing table listing cache read at $0.030 per 1M.

When the cheaper sibling wins

GLM-5.3-Flash wins when the job is throughput: bulk document classification, first-pass extraction, triage routing, anything where 18B active parameters and a 328 GB checkpoint are enough and the medical rubric is not the binding constraint. It also wins when the input is an image or a video, because LingSante cannot see either.

LingSante-3.1-flash wins when the job is clinical judgement on text: patient communication, evidence synthesis, structured record review, mental-health screening. The 65.72 HealthBench Professional figure and the 4.722 PatientAgentBench score are the strongest evidence in the release, even after discounting for vendor provenance.

The unresolved question is whether the 16-point medical gap survives the Flash's price advantage in a real pipeline. A two-stage design — cheap GLM for extraction, specialist for judgement — costs less than running the specialist on everything and may score better than either model alone. Nobody has published that comparison, and it is the one worth running before committing to a single vendor.

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.

So sánh trong bài viết này1

Phát hiện từ bài viết này · Benchmark: Artificial Analysis · cập nhật hằng ngày