Title card for a comparison of LingSante-3.1-air-VL and WeVisOne-VL 4B, with the subtitle Prose vs Pixels, cards reading Reads an image, writes findings and Reads a phrase, returns masks, and a footer noting both are open-weight with no hosted API.
Guides & Insights

LingSante-3.1-air-VL vs WeVisOne-VL-4B: Prose vs Pixels

Автор

jinhao.song

Дата публикации

Новые модели · 20Все модели →
Бенчмарки: Artificial Analysis · обновляется ежедневно
Назад ко всем статьям

On 22 November, Ant Ling posted a short thread introducing LingSante-3.1-air-VL — 124B total parameters, 5.5B active, open weights, medical. The checkpoint had actually been sitting on Hugging Face since 04:31 UTC on 19 November, uploaded with no announcement and no paper, and it still has zero downloads. T​encent's WeVisOne-VL 4B is the other end of the same idea: a 4.83B model released on 22 October that takes an image and a phrase and returns boxes, masks and alpha mattes. Put them side by side and the obvious question — which is better — is the wrong one, because the two models disagree about what a vision model is for.

Neither ships a hosted endpoint. Neither publishes a price. Both list a 262,144-token context and both are vendor-scored only. That is where the similarity ends, and the differences are the kind that decide an architecture rather than a leaderboard position.

Two models, two opposite output types

The cleanest way to separate them is to look at what comes out.

LingSante-3.1-air-VL reads pixels and writes language. Its stated jobs are medical image understanding, report generation, medical reasoning, multimodal search and medical GUI agents — every one of those ends in text, a structured finding, or a tool call. It is built on Ling-3.0-flash-VL, with a 27-layer vision tower (hidden size 1,152, patch 16, 2×2 spatial merging) feeding a 4,096-dimension projector into a 42-layer mixture-of-experts decoder with 512 routed experts and 8 active per token.

WeVisOne-VL 4B reads pixels and writes geometry. It emits coordinate tokens from <0> to <1000> directly in its vocabulary, plus an <OBJ> token whose hidden state feeds a separate 9.6M-parameter pixel decoder — a feature adapter, six deformable cross-attention blocks and a cascade head refining at strides 4, 2 and 1. That decoder turns predicted boxes into instance masks and full-resolution alpha mattes. There is no external segmenter in the loop.

The practical consequence is that the outputs are not interchangeable in either direction. You cannot take a segmentation mask out of the medical model, and you cannot take a radiology impression out of the perception model. A pipeline built for one will not accept the other's answer without a second model in between.

• Output — LingSante-3.1-air-VL returns text and tool calls vs WeVisOne-VL 4B returns boxes, points, masks and alpha mattes

• Domain — medical specialist, research-only vs general open-vocabulary perception and matting

• Base — Ling-3.0-flash-VL vs Qwen3-VL

• Licence — MIT with a research-only restriction vs Apache 2.0 for T​encent-authored code

Screenshot of the inclusionAI LingSante-3.1-air-VL model card on Hugging Face, showing 124B parameters and 5.5B active parameters per token, a 256K-token context window, the MIT licence tag, 8 likes, and a notice that the model is not deployed by any inference provider.Screenshot of the tencent/WeVisOne-VL-4B model card on Hugging Face, showing a 5B parameter Qwen3-VL model with task tags for object detection, referring expression comprehension, object counting, point grounding, OCR, referring expression segmentation and image matting, 36,254 downloads last month, 130 likes, and an Apache 2.0 licence tag.

The 262K context both claim, unpacked

On paper the two models tie at 262,144 tokens. In practice they got there in completely different ways, and only one of them can use it.

For LingSante-3.1-air-VL, the base config lists max_position_embeddings at 131,072, and the serving recipe extends it to 262,144 with YaRN factor 2.0. That is a documented server argument, not a benchmark claim, and it exists to make multi-image clinical reading work — a study with a stack of slices and a prior report is exactly the long-context case.

WeVisOne-VL 4B inherits 262,144 from Qwen3-VL natively, and the GGUF metadata confirms it. But the model is trained with up to 8,192 tokens of context, its evaluations cap images at roughly 1,960 visual tokens, and it accepts one image per turn while rejecting video grids outright. A 262K window on a single-image, single-turn model is capacity you pay KV-cache memory for and never fill: at full length the cache alone adds about 24 GB. Most quantized deployments sensibly run it at 16K to 128K.

• Advertised context — 262,144 vs 262,144

• How it was reached — YaRN 2.0 over a 131,072 base vs inherited natively from Qwen3-VL

• What fills it — multi-image clinical studies vs one image per turn, ~1,960 visual tokens

• Training context — not stated vs up to 8,192 tokens

What it costs to actually run each one

Neither model has a hosted API, so there is no per-token price to compare. Cost is hardware, and the gap there is roughly two orders of magnitude.

LingSante-3.1-air-VL is 249.7 GB of bfloat16 weights across 50 safetensors shards. The documented SGLang path uses tensor parallelism 4 on four 141 GB-class GPUs, or TP 8 on eight 80 GB H100/H800s, with the image lmsysorg/sglang:dev-Ling-3.0-flash-VL. The vLLM alternative needs inclusionAI's vllm-ling-v3 fork and --trust-remote-code. Either way you are renting a multi-GPU node before you score a single answer.

WeVisOne-VL 4B is about 19 GB in F32, 9.66 GB in BF16 and 2.9 GB at Q4_K_M. A third-party estimator puts it inside 8 GB of VRAM at Q4 with a 32K context, comfortably on a 24 GB card at 128K, and around 11 GB for the original weights under vLLM. Estimated generation throughput on an RTX 4090 is roughly 193 tokens per second — that figure is an estimate from llmrun.dev, not a measurement, and it should be treated with its stated ±20% band.

• Weights on disk — 249.7 GB bf16 vs 9.66 GB bf16 / 2.9 GB Q4_K_M

• Minimum hardware — four 141 GB-class GPUs (TP=4) vs a single 8 GB GPU at Q4

• Serving stack — SGLang or a custom vLLM fork vs Transformers, vLLM, SGLang, or the project package

• Published throughput — none vs ~193 tok/s on an RTX 4090 (third-party estimate)

This is also where a routing layer earns its place, for a mundane reason. Both of these are self-host projects sitting at the edge of a pipeline that still needs an ordinary text or triage model around them, and running that part through one API with a routing DSL means the surrounding steps do not acquire a second contract each time you swap a model. OrcaRouter does not list either LingSante-3.1-air-VL or WeVisOne-VL 4B today — both are open-weight downloads — but the failover in front of everything downstream is the part that keeps a 250 GB experiment from becoming a single point of failure.

Benchmarks with no shared row

There is no benchmark on which these two models both appear. LingSante-3.1-air-VL reports a 26-row scorecard of medical imaging and clinical reasoning tasks; WeVisOne-VL 4B reports detection, referring, counting, OCR, segmentation and matting. Every direct comparison would be an invention.

What the two scorecards do share is provenance. Both are vendor-run, and neither has an independent reproduction. Ant Ling's card shows 13 top-two finishes across 20 image benchmarks — 8 firsts and 5 seconds — with best values on CheXpert Plus (46.21 against GPT-6.1 Sol at 44.76), MIMIC-CXR (47.20 against 45.21), IU X-RAY (64.34), OmniMedVQA (93.64), VQA-RAD (81.15) and OmniDocBench v1.5 (91.20). Three of those rows are in-house AFUMED tasks that no outside team can reproduce.

T​encent's card claims LVIS F1 55.3 against a 50.7 best baseline, Dense200 F1 71.3 against 58.7, RefCOCOg test F1 81.8 against 77.6, ReasonSeg val gIoU 71.2 against 65.0, and P3M-500-NP matting MAD 7.3 against 15.4, where lower is better. The sharpest number is the within-family one: on Dense200, WeVisOne-VL 4B scores 71.3 while unmodified Qwen3-VL-4B scores 15.2.

The near-miss is document reading. LingSante's OmniDocBench v1.5 score of 91.20 and WeVisOne's TotalText F1 of 52.0 look like a head-to-head and are not: one measures document parsing against a page-layout benchmark, the other scene-text recognition under a custom mean over IoU 0.5–0.95. Different corpora, different metric, no comparison available. Directionally, though, it is the one row where their reputations invert — OCR is WeVisOne's weakest reported task and document reading is among LingSante's strongest.

A two-column scoreboard comparing LingSante-3.1-air-VL and WeVisOne-VL 4B across six dimensions: output, total parameters, weights on disk, context in practice, benchmark coverage and licence, with a footer stating all figures are vendor-reported with no independent benchmark for either model.

The medical question WeVisOne cannot answer

Anyone looking at a 124B medical model and a 4B general vision model will ask whether the small one can simply do the medical job. As of today there is no published evidence either way, and a structural reason to doubt it.

WeVisOne-VL 4B has not been evaluated on medical imagery in any source we could find. Its reported tasks are detection, dense detection, referring localization, rejection of absent targets, counting, OCR, referring segmentation and box matting — nothing clinical. More importantly, the fine-tuning code is still listed as unfinished, so you cannot adapt it to a clinical distribution yet even if you wanted to. The GGUF builds cannot produce masks at all, because the pixel decoder needs native ViT features.

The reverse assumption fails too. LingSante-3.1-air-VL is a medical specialist, but its text-only clinical reasoning is mediocre on its own scorecard: MedQA-USMLE 94.34 is last of eight, MedMCQA 81.54 is seventh, HealthBench Professional 49.59 is sixth behind its own 560B sibling at 65.72. Its strength is perception — reading an image and producing findings — not answering a clinical knowledge question in the abstract.

So the honest split is narrower than the names suggest. The 124B model is a medical perception engine with weak general clinical reasoning; the 4B model is a strong general perception engine with no medical evidence and no adaptation path yet. Neither is a diagnostic system, and both say so: LingSante's card restricts it to research and states it is not a medical device, while WeVisOne's intended use excludes autonomous high-impact decisions.

Failure modes that decide a deployment

Where independent testing is missing, the vendor's own limitation notes are the best available signal, and both cards are unusually candid.

WeVisOne-VL 4B's rejection behaviour varies sharply by benchmark. WeRef-Bench rejection is 96.4%, but PR-Bench is 52.6% and RefCOCOg-Rej is 67.4% — meaning on some absent-target queries roughly one in three still returns a fabricated box. Prompt wording changes the output: "that woman's hair" and "her hair" are not guaranteed to produce the same alpha, and detection requires one label per line or the model reads the list as a single phrase and finds nothing. There is no chain-of-thought, so an ambiguous instruction is a one-shot bet. Alpha quality degrades on fine boundaries, heavy occlusion and transparent materials.

LingSante-3.1-air-VL's weaknesses are mostly about maturity rather than behaviour. Zero downloads and eight likes as of this writing means no community validation of any kind; the repository was still receiving commits on 22 November; and the benchmark card mixes reproducible public tasks with in-house ones. The research-only restriction is explicit, and the medical framing carries obvious risk if anyone reads it as clinical software.

• Absent-target behaviour — n/a, not a detection model vs 52.6%–96.4% rejection depending on benchmark

• Prompt sensitivity — not documented vs high, mask changes with phrasing

• Community validation — none (zero downloads) vs 36,254 downloads, community GGUF and LiteRT ports

• Independent benchmark — none vs none

Кто должен выбирать

The decision reduces to the output type, then to hardware.

Pick LingSante-3.1-air-VL if you already operate datacenter GPUs, your task ends in text about an image — structured findings, report drafts, medical VQA, multi-image study review — and you are prepared to validate it on held-out data because nobody else has. Budget for four 141 GB-class GPUs and accept that the numbers you are planning against are vendor-run.

Pick WeVisOne-VL 4B if you need geometry: boxes, points, instance masks or alpha mattes, at a scale where a single 24 GB card is the whole deployment. It is the only one of the two that returns pixels, it runs on hardware you probably already own, and it has a month of community tooling behind it in the form of GGUF and LiteRT ports and a ComfyUI integration.

Pick neither if you need a production diagnostic path this quarter. One model has no independent evaluation and no downloads; the other has no medical evidence and cannot yet be fine-tuned. That is not a knock on either — it is what four days and thirty-two days of public life buys you.

Вопросы, которые стоит задать, прежде чем решиться

Can WeVisOne-VL 4B read a chest X-ray?

It can describe what it sees, because its Qwen3-VL base has general visual competence, but no published evaluation tests it on radiology and its reported tasks are detection, OCR, segmentation and matting. The more relevant limit is that fine-tuning code is unreleased, so a team cannot correct its behaviour on clinical images yet. Treat any medical output from it as an untested general-model answer, not a measurement.

Is LingSante-3.1-air-VL cheaper to serve than WeVisOne-VL 4B?

Neither has a published per-token price, so the comparison is hardware cost and it is not close. The medical model needs 249.7 GB of weights and a four-to-eight-GPU node; the perception model needs 2.9 GB at Q4 and runs inside 8 GB of VRAM. At any realistic utilisation the 4B is one to two orders of magnitude cheaper per image, and the 124B model has to justify that gap with capability you cannot get elsewhere.

Does either model need a second segmentation model bolted on?

WeVisOne-VL 4B does not — its 9.6M-parameter pixel decoder produces masks and alpha mattes from the model's own coordinate tokens, which is the entire point of the release. LingSante-3.1-air-VL does not produce masks at all, so if you need segmentation alongside medical findings you are adding a separate model either way. That asymmetry, more than the parameter count, is what decides which of these belongs in a given pipeline.

The two releases are usually filed under the same heading — open-weight multimodal models — and that filing is what makes the comparison misleading. One is a 124B medical reader that turns a scan into a paragraph and needs a datacenter to do it. The other is a 4.83B perception engine that turns a phrase into a mask and runs on a consumer card. Only one of them can give you pixels, only one of them can give you findings, and neither has been checked by anybody outside the lab that built it.