LFM3-VL-2B vs North Micro Vision Instruct: Two Document Readers, One Licence Wall

The accuracy gap between LFM3-VL-2B and North Micro Vision Instruct is real but small: about 4.6 points on DocVQA and 7.1 on OCRBench, both in the same direction. The gap that decides deployments is not a benchmark at all. North Micro Vision Instruct ships under Apache 2.0. LFM3-VL-2B ships under the LFM Open License v1.0, which licenses commercial use only while your entity stays below $10 million in annual revenue.

Both models are small, both are open-weight, and both were built to read pages rather than to chat about photographs. One was published in August 2026 and has had two months of downloads and ports. The other appeared on Hugging Face on 19 October, was formally announced on 27 October, and has 255 downloads. This comparison is therefore partly a benchmark question and partly a question about which risk you can live with: an unproven model, or an unlicensed one.

What each model actually is

LFM3-VL-2B is Liquid AI's 1.92B dense model: a 1.73B language backbone, a 164M in-house vision encoder built to use roughly three times less compute than SigLIP2, and a 25M projector. It takes text, images and video at 2 fps, outputs text, supports a thinking mode, carries a 128K context and covers 16 languages. The repository includes BF16, GGUF, MLX 4-bit and ONNX builds, though llama.cpp support is still in review and has no video path.

North Micro Vision Instruct is Cohere Labs' 2.4B model: a 2B language backbone from the North Micro family plus a roughly 400M encoder continued from SigLIP2 SO400M, with a DeepStack-style projector. It takes interleaved text and images at native resolution, up to 1654×2339 — a full A4 page at 200 dpi — and outputs text. The language model has a 128K context, but Cohere states that multimodal training was validated only to 8K. Apache 2.0, 262K vocabulary, 11 to 12 languages.

Those specifications look similar and are not. A 128K context is a ceiling on tokens the model can attend to. Native-resolution input is a decision not to resize the page in the first place. For a scanned invoice with a table in the corner, the second matters more than the first.

The document numbers, read with the right reservations

On the benchmarks both cards publish, LFM3-VL-2B leads:

• DocVQA — LFM3-VL-2B 96.7 vs North Micro Vision Instruct 92.1.

• OCRBench — 86.3 vs 79.2.

• ChartQA — 92.7 vs 80.8.

• AI2D — North reports 77.5; Liquid does not publish a directly comparable AI2D row.

• RefCOCO-avg — 90.5 vs 73.2, though the two cards describe these averages differently enough that the subtraction is indicative rather than exact.

All of these are vendor-reported, on harnesses the vendors chose, with no independent reproduction of either model. The DocVQA gap is narrow enough that a prompt-format difference could close it. The ChartQA gap, at nearly 12 points, is less likely to be noise, and the OCRBench gap of 7.1 points is consistent with a model that was explicitly trained on document and chart data with a reasoning mode attached.

The interesting number is not on the document rows. It is MMMU, where LFM3-VL-2B reports 55.7 and North Micro Vision Instruct reports 32.9. That is not a tuning difference; it is a different product. Cohere's own card says the model is deliberately not a reasoning model, does not do tool calling, and discourages system prompts. It reads pages. Liquid's card claims graduate-level reasoning as one of 62 measured categories. If your task is "extract the totals from these 400 invoices," both work. If your task is "decide whether this invoice violates policy," only one was built for it.

Where the 128K claim gets slippery

LFM3-VL-2B's 128K context is the more generous published figure, but context length is not the same as usable multimodal context. The card does not state how many image tokens a page consumes at native resolution, and it publishes no maximum frame count for video. North Micro Vision Instruct publishes the opposite kind of number: a 128K language context with multimodal training validated to 8K. That is a conservative disclosure, and it is more useful than a larger number with no stated vision budget.

The practical test for either model is the same: measure tokens per page on your own documents before you design around a context window. A 128K window that holds four pages at your resolution is not a 128K document workflow.

The licence is the decision

Apache 2.0 is Apache 2.0. You can use North Micro Vision Instruct commercially at any revenue, modify it, redistribute it and host it, provided you keep the notices. That is the whole agreement, and for a startup building a document product it removes an entire category of legal review.

The LFM Open License v1.0 is not that. It defines commercial use broadly — "any use of the Work for direct or indirect commercial advantage or monetary compensation" — and conditions the licence on the licensee staying below a $10 million annual revenue threshold. Above that threshold, commercial use "is not licensed under this Agreement." Qualified non-profits are exempt for non-commercial and research purposes. Redistribution is allowed with attribution and change notices. The licence text we read contains no separate hosting restriction, so the revenue ceiling, not a serving ban, is the constraint.

That ceiling produces an awkward conversation. A company at $9 million in revenue can ship LFM3-VL-2B today and lose the licence by growing. The model does not stop working; the permission does. For a product with a two-year roadmap, that is a diligence item, not a footnote.

The adoption gap is its own signal

North Micro Vision Instruct has been downloaded about 35,000 times since its repository appeared on 10 August 2026. LFM3-VL-2B has 255 downloads and 34 likes since 19 October. Two months versus nine days explains part of that, but not all of it: the LFM3 family announcement on 27 October was a single blog post covering six models, and the vision checkpoint got one paragraph of it.

Quiet releases are not automatically bad releases. A model with no launch pricing, no hosted provider and framework support still in review is a model you evaluate on your own hardware, and the low download count mostly means nobody has done that yet. It also means there is no community quantisation, no issue tracker full of fixes and no third-party benchmark to lean on. North Micro Vision has all three.

What each one costs, as far as anyone will say

Neither model publishes an API price on its card. North Micro Vision Instruct is available through Cohere's platform and has been picked up by several third-party hosts; LFM3-VL-2B is available on Liquid's Playground and API with no rate card, and Hugging Face lists no inference provider for it. On owned hardware the comparison is 2.4B parameters against 1.92B, or roughly 4.8 GB against 3.84 GB in BF16 — close enough that GPU cost is not the tiebreaker.

The tiebreaker is operational. If you rent a hosted vision model instead of owning either checkpoint, OrcaRouter passes provider list price through at 0% markup, so a vendor price change is live on our side the same day rather than at the next contract renewal. We do not serve LFM3-VL-2B or North Micro Vision Instruct today, and this piece does not claim otherwise; the point is that the hosted alternative is priced transparently enough to compare against the self-hosting arithmetic above.

Who should pick which

Choose North Micro Vision Instruct if you need Apache 2.0, if you want a model with two months of community use behind it, or if your documents are full pages at native resolution and your task stops at extraction. Accept that it will not reason across the extracted fields.

Choose LFM3-VL-2B if you need video input, a thinking mode, or a broader visual assistant that happens to read documents well, and if you are comfortably below the revenue threshold or can negotiate a separate licence. Accept that you are an early adopter of a checkpoint with no independent evaluation and incomplete framework support.

Choose neither, yet, if your workflow needs a priced hosted endpoint with a support contract. The small open-weight tier is improving quickly, but at 255 downloads the newest entrant has not been stress-tested by anyone outside its lab.