
LingSante-3.1-air-VL vs LFM2.5-VL-3B: 3 GB Against 250 GB
- OrcaNYOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 186 tok/s
- openaiNYOpenAI: GPT-6.2 Astra2026-11-17$8.00 / $40.00 per 1M tokens · 87 tok/s
- openaiNYOpenAI: GPT-6.2 Sol2026-11-17$2.00 / $10.00 per 1M tokens · 151 tok/s
- tencentNYTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 187 tok/s
- AlibabaNYQwen: Qwen4 Max2026-11-16$1.50 / $4.50 per 1M tokens · 85 tok/s
- GoogleNYGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 705 tok/s
- AnthropicNYAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 239 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 96 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 138 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 85 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 122 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 101 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 70 tok/s
- openaiOpenAI: GPT-6.1 Terra2026-10-26$1.00 / $5.00 per 1M tokens · 35 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 89 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 38 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligens
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 231 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligens
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 101 tok/s
The two models in this comparison are separated by a factor of roughly seventy-five in memory. LFM2.5-VL-3B runs in about 3.3 GB and answers at 228 tokens per second on a laptop. LingSante-3.1-air-VL needs 249.7 GB of weights and four 141 GB-class GPUs before it produces a token. Put them side by side on a spec sheet and the exercise looks absurd — a 3.1-billion-parameter edge vision model from Liquid AI against a 124-billion-parameter medical mixture-of-experts that Ant Group uploaded without announcing.
It is not absurd, because the people searching this comparison are not asking which model is better. They are asking whether a small on-device model is now good enough for clinical document and image work, or whether that job still requires a datacentre. The honest answer turns on something the benchmark tables cannot supply: the two cards share no comparable rows, so any "X beats Y" verdict would be invented. What can be compared is the job each model was built to do, the hardware each one demands, and the licence each one carries.
Two models, two very different jobs
LFM2.5-VL-3B is an edge model with a release date. Liquid AI published it on 12 August 2026, and the ecosystem has had three months to work on it: 62,474 downloads, 239 likes, official GGUF, ONNX, MLX and LiteRT builds, and a separate speculative-decoding drafter. Its card is explicit that it is a non-reasoning model — it answers directly rather than thinking first — and that it is not recommended for reasoning-intensive work such as visual web design or technical questions about blueprints.
LingSante-3.1-air-VL is a medical specialist with no release date at all. inclusionAI created the repository at 04:31 UTC on 19 November 2026, published an MIT licence, and said nothing. A couple of likes, zero downloads, no blog entry, no hosted endpoint. Its card describes four clinical focus areas: imaging reports, multimodal health assessment covering skin, diet and reports, medical search and evidence synthesis, and medical GUI agents. Nothing on the card suggests it should run anywhere except a GPU cluster.
The spec sheet, without the marketing
• Parameters — 3.1B total for LFM2.5-VL-3B, built from a 2.6B language backbone and a 400M SigLIP2 NaFlex vision encoder, versus 124.8B total and 5.5B active per token for LingSante-3.1-air-VL.
• Memory — 6.26 GB of repository weights and under 3.3 GB at runtime for the 3B; 249.7 GB of bfloat16 weights in fifty shards for the medical model.
• Context — 32,768 tokens for LFM2.5-VL-3B; 262,144 for LingSante-3.1-air-VL, enabled by a published YaRN factor of 2.0 over a 131,072-token base.
• Modalities — text and images on both. The Ling config also reserves video token IDs, though its card claims only text and images.
• Languages — sixteen for LFM2.5-VL-3B, spanning Arabic, Chinese, Hindi, Japanese, Korean, Thai, Vietnamese and most of Europe; the Ling card makes no language claim.
• Serving — llama.cpp, MLX, vLLM, SGLang and ONNX for the 3B. SGLang with a day-zero image at tensor parallelism 4, or a custom vLLM fork requiring --trust-remote-code, for the medical model.
• Licence — LFM Open License v1.0 for LFM2.5-VL-3B, a custom licence rather than an OSI one, against plain MIT for LingSante-3.1-air-VL.
• Maturity — three months of public use and four quantisation formats against four days, a couple of likes and no third-party build of any kind.
No shared benchmark rows, by design

This is the part that makes a score comparison dishonest, so it is worth showing the mechanics rather than asserting the conclusion.
LFM2.5-VL-3B reports ScreenSpot-v2 at 80.7, RefCOCO at 87.9, BLINK at 61.5, MuirBench at 58.3, DocVQA at 91.1, ChartQA at 81.3, OCRBench v2 at 47.5, MMMU at 48.4 and MMMU-Pro at 30.5, among twenty-eight tasks with a reported average of 69.4. It is benchmarked against Gemma 4 E2B and E4B, InternVL 3.5 at 2B and 4B, and Qwen3.5 at 2B and 4B — models in its own weight class, which is the correct comparison.
LingSante-3.1-air-VL reports none of those tests. Its twenty-six rows are medical: CheXpert Plus, IU X-RAY, MIMIC-CXR, OmniMedVQA, PathVQA, PMC-VQA, SLAKE, VQA-RAD, GMAI-MMBench, MedQ-Bench, MedQA, MedMCQA, HealthBench Professional, OmniDocBench, BrowseComp-VL, MMSearch, AndroidWorld and three in-house AFUMED sets. Its comparators are GLM-5.3-Flash, DeepSeek-V4.1-Flash, MiniMax-M3.1-Flash, Kimi-K3.1-Pro and three frontier closed models.
Even the one pair that looks comparable is not. LFM2.5-VL-3B scores 30.5 on MMMU-Pro. LingSante-3.1-air-VL scores 73.95 on MMMU-Pro-Med, a medical subset with a different question distribution. Setting those numbers beside each other would produce a forty-three-point gap that measures the subset, not the models. A reader who wants a number to quote here should accept that no number exists yet.
What a 3B model can actually do here
The 3B's strengths are narrower than "vision-language model" suggests, and more useful than the parameter count suggests. Its release notes dwell on three things: grounding, OCR and speed.
Grounding means natural-language object detection and referring expressions — RefCOCO at 87.9 macro precision, up from 57.1 in the previous generation. In a clinical intake pipeline that is the difference between "there is text in this scan" and "the patient name is in this box." Its OCR path is more specific still: full-page OCR with layout annotation, emitting labelled regions with bounding boxes in a documented format. For turning a stack of scanned referrals or lab printouts into searchable text with structure, that is the exact capability required.
Speed is where it becomes a design option rather than a toy. Liquid's figures are 228 tokens per second on an Apple M5 Max, 116 on an AMD Ryzen AI Max+ 395, about 20 on a Galaxy S26 Ultra, and roughly 11,000 output tokens per second on a single H100 at high concurrency. At 3.3 GB it fits on hardware a clinic already owns, which changes the privacy argument: if the image never leaves the building, the compliance conversation is shorter.
The limits are equally concrete. A 32,768-token window cannot hold a patient record plus a study plus a guideline. Non-reasoning output is fast and shallow by construction. And the card itself steers you away from technical, reasoning-heavy questions — which, unfortunately, is most of clinical interpretation.
What only the 124B model can do

The medical model's card is written around tasks the 3B was not trained for: reading multiple images from one study together, weighing patient history and imaging technique before reviewing the images, and writing structured Findings and Impression sections in clinical language. It claims the model can identify eight or nine types of positive finding on a single complex chest radiograph — cardiomegaly, pleural effusion, atelectasis, consolidation, pneumonia, pneumothorax — and can ask follow-up questions across turns before assessing a skin lesion, drawing on a stated 120-plus skin conditions.
Its scorecard backs the perception half of that claim, at least on the vendor's own numbers. First place on CheXpert Plus at 46.21, MIMIC-CXR at 47.20, IU X-RAY at 64.34, OmniMedVQA at 93.64 and VQA-RAD at 81.15. Second place on PathVQA, PMC-VQA and SLAKE, each time behind Claude Opus 5.5 rather than behind a small model. It also carries document parsing at 91.20 on OmniDocBench v1.5 and GUI-agent results on AndroidWorld at 86.84.
Where it does not back the claim is anything requiring medical knowledge without an image. HealthBench Professional 49.59. MedQA-USMLE 94.34, last of eight. MedXpertQA-Text 50.65. The model sees well and reasons thinly, which is the opposite failure mode from a cheap edge model that sees adequately and reasons not at all.
A staged pipeline beats either alone
Once the comparison is framed as two different jobs, the useful architecture is not a choice. It is a queue.
Stage one runs on hardware you own: capture, deskew, OCR with layout, classify the document type, detect whether the image is even a medical image, and reject the obvious failures before they cost GPU-seconds. LFM2.5-VL-3B is well suited to that stage and its licence permits commercial deployment under Liquid's terms.
Stage two runs where the images are already meant to be processed: multi-image reading, structured findings, report drafting, and evidence lookup against guidelines. That is the stage LingSante-3.1-air-VL was trained for, and it is the stage that needs 250 GB of accelerator memory.
Routing is the part that usually gets hand-built badly. OrcaRouter exposes 200-plus models behind a single key at provider list price with 0% markup and automatic failover, and its routing DSL composes several models into one call — the vision stage and the text stage can come from different vendors without a second contract or a rewrite when one of them changes. Neither of these two models is in the catalogue today; the relevance is structural, because a two-stage clinical pipeline is exactly the shape that layer is for.
There is a cheaper version of the same idea worth testing first. Run the 3B on a sample of your own documents, score the output, and only then ask whether stage two is needed. If the work is intake and retrieval, it may not be. If the work is a draft that a clinician signs, it almost certainly is.
Cost, licence and the datacentre question
Neither model has a per-token price, and that is the quiet trap in this comparison. A reader who expects two API price columns will not find them. LFM2.5-VL-3B is open weights with no published endpoint pricing; LingSante-3.1-air-VL has no hosted endpoint at all. So the cost question is entirely a hardware question.
• Floor — a laptop or a single workstation GPU for the 3B at 3.3 GB; four 141 GB-class GPUs, or eight 80 GB H100/H800s, for the medical model.
• Throughput — thousands of output tokens per second per H100 for the 3B; no published throughput figure for the medical model, and no host to measure it on.
• Quantisation — official GGUF, ONNX, MLX and LiteRT builds for the 3B; none for the medical model, four days after upload.
• Licence — LFM Open License v1.0 with commercial terms attached, against MIT with none.
• Verification — both scorecards are vendor-reported. The difference is that 62,000 downloads have given the 3B nine months of community scrutiny, while the medical model has had none.
That last line is the one to hold onto. The 3B's benchmark claims have been argued about in public, including by critics who note that a 3B-class average flatters a model against the weak spots of 4B rivals. The medical model's claims have not been tested by anyone outside Ant Group.
Verdict: pick the tier, not the winner
If the work is document intake, layout-aware OCR, grounding, or anything that must run where the data already lives, LFM2.5-VL-3B is the operating choice today. It is small, fast, licensable for commercial use, three months into community validation, and available in four deployment formats. It will not write an imaging report and its card does not pretend it will.
If the work is reading studies, drafting findings, or reasoning across several images at once, LingSante-3.1-air-VL is the only one of the two built for it — and it is not yet a safe production dependency. No independent benchmark, no hosted endpoint, no quantisation, no announcement, four days old. The responsible move is to download it, run it on your own held-out set, and treat the result as a pilot, with a fallback model behind it.
The question in the title has a real answer, and it is not the one the parameter counts suggest: the 3 GB model is the one you can deploy this quarter, and the 250 GB model is the one worth evaluating for next year.

