A generated title card reading 'LingSante-3.1-air-VL vs North Micro Vision Instruct' with the subtitle 'A fine-tuning base against a finished specialist'. Three icon cards read '2.4B dense against 124B MoE', '14 quantisations against none' and 'Apache-2.0 against MIT'. A footer reads 'Specs from both model cards, 23 November 2026.' The OrcaRouter logo is composited in a white strip along the bottom edge.
Guides & Insights

LingSante-3.1-air-VL vs North Micro Vision Instruct: Build or Buy

Autor

jinhao.song

Data publikacji

Najnowsze modele · 20Zobacz wszystkie modele →
Benchmarki: Artificial Analysis · aktualizowane codziennie
Powrót do wszystkich wpisów

North Micro Vision Instruct is a 2.4-billion-parameter model that Cohere Labs released on 10 August 2026 under Apache-2.0, and its release has the texture of something meant to be taken apart: native-resolution image handling that preserves aspect ratios, eleven languages, multiple images per conversation, and fourteen quantised builds listed on the Hub. LingSante-3.1-air-VL is a 124-billion-parameter medical mixture-of-experts that inclusionAI uploaded on 19 November 2026 under MIT, and it has the texture of something meant to be used as shipped: a finished clinical reading model with no quantisations, no fine-tuning recipe, and a serving command that assumes four 141 GB accelerators.

Those are not competing answers to the same question. They are the two ends of a decision every clinical-document team eventually faces: do you take a small general model and teach it your task, or do you take a large specialist and adapt your workflow to what it already does? The parameter counts get the headlines. The fine-tuning surface is what actually determines which one you can live with.

A 2.4B model with fourteen builds

Screenshot of the Hugging Face model card for CohereLabs/North-Micro-Vision-Instruct captured 23 November 2026 in English: 155 likes, Apache-2.0 and 11 languages tags; sidebar showing 16,649 downloads last month, 'Model size 2B params' in BF16, the notice 'This model isn't deployed by any Inference Provider', a model tree listing 1 adapter, 5 finetunes, 1 merge and 14 quantizations, one Space, and a linked Cohere article dated 12 August; body text describing a 2.4B-parameter open-weight vision-language model with native-resolution image support, built as a foundation for prototyping and task-specific fine-tuning, with highlights for aspect-ratio-preserving images, VQA, captioning, grounding, OCR, charts and documents, eleven-language and multi-image support.

The North Micro Vision card reads like a foundation, and the ecosystem has treated it that way. Since August the MLX community has published 4-bit, 5-bit, 6-bit and 8-bit conversions alongside mxfp8, mxfp4, nvfp4 and bf16 builds, with fourteen quantised checkpoints now listed under the model on the Hub — which is what a healthy small-model release looks like after a few months. The repository has drawn roughly 16,600 downloads in the past month and 155 likes.

The defining feature is native resolution. Rather than resampling every image into a fixed grid, the model preserves aspect ratio and fine visual detail, which matters for scanned forms, lab printouts and screenshots where a resampled glyph is a lost character. The card lists VQA, captioning, grounding, OCR, charts and documents as the intended surface, and eleven languages: English, German, French, Spanish, Italian, Portuguese, Hindi, Japanese, Korean, Chinese and Arabic. Weights are 2,484,847,856 parameters in bfloat16.

Adoption is evidence of usability, not accuracy. People convert models that are easy to convert, and a clean 2.4B Apache-2.0 release is easy. Cohere's quality claims are Cohere's own, and no neutral leaderboard entry has settled them. What is not in doubt is the shape: small, dense, every parameter active, and cheap enough to run beside the application rather than behind an API.

What each one is designed to be

The design briefs diverge before any benchmark is involved.

• Parameters — North Micro: 2.48B dense. LingSante: 124B total, 5.5B active per token, 512 routed experts in a 42-layer decoder.

• Image handling — North Micro: native resolution, aspect-ratio preserving. LingSante: a 27-layer ViT with patch size 16 and 2×2 spatial merging, plus a 4,096-dimension projector into the language model.

• Context — North Micro: sized for its 2.4B deployment class. LingSante: 262,144 tokens, reached by a documented YaRN factor of 2.0 over a 131,072 base.

• Footprint — North Micro: a small BF16 checkpoint that fits on a workstation, with quantised builds down to 4-bit. LingSante: 249.7 GB of bfloat16 across fifty shards, needing tensor parallelism across four 141 GB GPUs or eight 80 GB cards.

• Licence — Apache-2.0 for North Micro, MIT for LingSante, with a research-only, not-a-medical-device notice added in the medical model's prose.

• Stated purpose — North Micro: prototyping, task-specific fine-tuning, specialised multimodal work. LingSante: medical imaging reports, skin and diet assessment, report interpretation, medical search and GUI agents.

One of these is a component. The other is a product with a narrow job description. That distinction survives every benchmark argument, because it is about who does the adaptation.

Native resolution against 256K context

The two engineering choices answer different failure modes. Native resolution is a bet that detail is lost in preprocessing, and that a model which sees the original pixels will read a lab value, a dosage or a handwritten note better than one that sees a downsampled tile. It is a sensible bet for enterprise document work, where the input is a scan and the answer is a field extraction.

LingSante's 256K context is a bet that the problem is the record, not the pixel. The card describes a model that takes patient history, imaging technique and image quality into account before reviewing images systematically, then writes Findings and Impression. It can read multiple images together and identify eight or nine positive findings on one complex chest radiograph. That is a claim about assembling context across images and text, and it is the kind of claim a 2.4B model does not make.

Both bets are vendor-stated. Neither has been independently measured. The difference is that one can be tested on a laptop this afternoon and the other needs a datacentre allocation before the first token is generated.

The document question neither answers

Here is where the comparison gets uncomfortable for both cards. A clinical document pipeline usually has two stages: read the page, then decide what the page means. North Micro addresses the first and explicitly leaves the second to your fine-tune. LingSante addresses both, but its own scorecard shows the reasoning half is thin — MedQA-USMLE 94.34 is last of eight comparators, HealthBench Professional 49.59 is sixth, and its 560B sibling scores 65.72 on the same row.

LingSante does publish one document number: OmniDocBench v1.5 at 91.20, ahead of Kimi-K3.1-Pro at 91.05 and Claude Opus 5.5 at 90.56. That is a strong row and it is vendor-reported on a single benchmark. North Micro's card claims OCR, chart and document capability but does not give us a comparable per-benchmark figure, so setting the two side by side would invent a gap rather than measure one.

What can be said is that a 2.4B model with native resolution and a fine-tuning path is the more predictable choice for a fixed extraction schema — invoices, requisitions, lab headers — because you can train it on your own labelled pages and measure the result yourself. LingSante is the more interesting choice for prose-heavy interpretation, where you cannot easily write the label set, and correspondingly harder to validate.

Fine-tuning is the real fork

Everything above collapses into one practical question: can you afford to own the adaptation?

North Micro ships as a base. Apache-2.0, 2.4B dense, quantised builds already circulating, designed for task-specific fine-tuning — a team with a few thousand labelled pages and one workstation can produce a specialised extractor in weeks. The model becomes yours in a way a hosted API never does, and the cost curve is predictable.

LingSante does not ship as a base in any practical sense. There is no fine-tuning recipe, no LoRA configuration, no quantisation, and no community tooling yet. Full-parameter training on 124B is a different budget category, and even LoRA work requires holding the checkpoint somewhere. inclusionAI built the medical capability in and expects you to use it as delivered. That is a legitimate model of release. It is also a much longer commitment than downloading a 5 GB checkpoint.

The middle path is to stop treating this as either/or. A small fine-tuned reader extracts structure; a large specialist handles the interpretation that resists a label schema. That is two models with two jobs, and it is exactly the shape a router exists to serve. OrcaRouter's model fusion lets a panel of models answer the same call together, which is a cheap way to cross-check an extraction before it reaches a clinical record; automatic failover keeps the pipeline alive if one provider degrades. Neither North Micro Vision Instruct nor LingSante-3.1-air-VL is in the OrcaRouter catalogue today — the value here is in composing whatever you do deploy, not in pretending we host these two.

A decision, not a winner

A generated two-column comparison scoreboard titled 'LingSante-3.1-air-VL vs North Micro Vision Instruct — the scoreboard'. The left column, LingSante-3.1-air-VL, reads 'Parameters: 124B total / 5.5B active', 'Weights on disk: 249.7 GB, 50 shards', 'Quantised builds: none published', 'Fine-tuning path: no recipe', 'Licence: MIT' and 'Document score: OmniDocBench 91.20'. The right column, North Micro Vision Instruct, reads 'Parameters: 2.48B dense', 'Weights on disk: one BF16 checkpoint', 'Quantised builds: 14 on the Hub', 'Fine-tuning path: the design brief', 'Licence: Apache-2.0' and 'Document score: no public figure'. A footer reads 'Vendor-reported; the two cards share no benchmark, 23 November 2026.' The OrcaRouter logo is composited in a white strip along the bottom edge.

Choose North Micro Vision Instruct if the job is document extraction with a schema you control, the hardware is a workstation, and the team has labelled data and the appetite to fine-tune. Its native-resolution handling and quantised builds make it the lower-risk engineering choice, and its adoption suggests the tooling will keep improving.

Choose LingSante-3.1-air-VL if the job is clinical interpretation, the hardware already exists, and you want a model that reads images and writes a structured report without you building that capability from scratch. Budget for evaluation rather than training, and treat the research-only notice as binding.

Choose neither, yet, if what you actually need is a general document assistant. North Micro is underpowered for open-ended reasoning and LingSante is overpriced for extraction. The release that would change this is a LingSante checkpoint with a published fine-tuning path, or a North Micro successor that ships per-benchmark document scores. Neither exists today, and the absence is the most useful fact on either card.

Screenshot of the Hugging Face Files and versions page for inclusionAI/LingSante-3.1-air-VL captured 23 November 2026 in English: the main branch is labelled 'LingSante-3.1-air-VL' at 250 GB with 12 commits and 4 contributors; the newest commit reads 'Update README.md with local benchmark assets and latest model card' and is marked as recent; the file list shows an assets folder, README.md at 9.1 kB, config.json, chat_template.jinja, bailingmm Python modules, and safetensors shards numbered model-00001-of-00050 through model-00014-of-00050, each 5 GB, committed two days earlier.

Porównane w tym artykule1

Wykryto na podstawie tego artykułu · Benchmarki: Artificial Analysis · aktualizowane codziennie