LFM3-VL-2B vs MiniCPM-V 4.7 1B: The 14-Point Gap That Shrinks to 4 When You Turn Thinking Off

The most consequential difference between LFM3-VL-2B and MiniCPM-V 4.7 1B is not in either benchmark table. It is in the licence line. OpenBMB's model is Apache 2.0. Liquid AI's is the LFM Open License v1.0, which carries commercial conditions instead of plain permissive terms. The two models are otherwise close enough in size to be direct substitutes — 1.92B against 1.55B — and they arrived within three weeks of each other: MiniCPM-V 4.7 1B went up on Hugging Face on October 6 and was dated October 12 in the vendor's release notes, while LFM3-VL-2B sat in a public repository from October 19 and was announced on October 27. On Liquid's own 62-benchmark suite the new model averages 71.4 against 57.5, a 13.9-point gap that collapses to 4.6 points the moment you disable reasoning on the Liquid model.

Two models in the same weight class, under different rules

MiniCPM-V 4.7 1B is the smaller of the two by about 370 million parameters. It combines a SigLIP-style vision tower with a Qwen3.5-derived language backbone, uses 4× or 16× visual token compression selectable per request, and claims a 256K context window in the vendor's documentation. It ships in BF16, int4, AWQ, GPTQ and GGUF builds, with a dedicated DSpark draft model for speculative decoding.

LFM3-VL-2B pairs the LFM3-1.7B language model with a new in-house 164M vision encoder and a 25M projector, at a documented 128K context, with video input at 2 frames per second and a documented no-thinking mode.

• Parameters — 1.92B for LFM3-VL-2B vs 1.55B for MiniCPM-V 4.7 1B, per OpenBMB's documentation; Hugging Face's model-size badge rounds the latter to 2B.

• Licence — LFM Open License v1.0 vs Apache 2.0.

• Context — 128K tokens vs a claimed 256K, though inference configurations commonly set the ceiling far lower.

• Vision path — a 164M in-house encoder with one token per 2×2 patch block vs a SigLIP-style tower with 4×/16× compression.

• Video — 2 frames per second, first LFM vision model to accept video vs supported, with LongVideoBench and MotionBench figures published by the vendor.

• Reasoning mode — available and measured both ways vs available and used in the benchmark run.

• Vendor average on Liquid's 62-benchmark suite — 71.4 thinking, 62.1 no-thinking, vs 57.5.

The aggregate gap is 13.9 points. Read the footnote first.

Liquid's table is the only artefact that scores both models on the same 62 benchmarks with the same normalisation, and it is the reason this comparison can be made at all. OpenBMB's own card does not benchmark against LFM3-VL-2B, which did not exist when MiniCPM-V 4.7 1B shipped.

The footnote that matters: reasoning mode was used for every model in the table except LFM2.5-VL-3B, which does not support it, and the LFM3-VL-2B no-thinking configuration. So 71.4 and 57.5 are both reasoning-enabled scores, which makes the comparison fairer than the LFM2.5 matchup. Both models are being measured at their best setting, on the same instruments, by the same lab.

Fair is not the same as independent. Liquid chose the benchmarks and ran them. The table is unusually complete — it includes category-level and per-benchmark rows, and it names every competitor — but no third party has reproduced any of it.

Turn thinking off and the 1B wins 24 of 62

LFM3-VL-2B publishes a second score for the same weights with reasoning disabled. That column is the interesting one, because it shows where the 13.9-point aggregate actually comes from. Without thinking, LFM3-VL-2B wins 35 benchmarks, MiniCPM-V 4.7 1B wins 24, and 3 tie. The average gap falls to 4.6 points.

More striking is what the 1B wins in that configuration. It takes most of the mathematical and chart-reasoning cluster, often by a wide margin.

• We-Math (testmini, strict) — 47.8 for MiniCPM-V 4.7 1B vs 20.1 for LFM3-VL-2B without thinking, and 68.6 with thinking.

• LogicVista — 46.1 vs 30.2 no-thinking, 60.4 with thinking.

• MathVista (mini) — 71.2 vs 68.5 no-thinking, 84.2 with thinking.

• ChartQA (test) — 88.8 vs 81.4 no-thinking, 92.7 with thinking.

• CharXiv (descriptive) — 73.4 vs 65.5 no-thinking, 87.5 with thinking.

• DynaMath — 63.8 vs 57.7 no-thinking.

• MMMU (val) — 51.0 vs 48.6 no-thinking, 55.7 with thinking.

The We-Math result is the one worth pausing on: a 27.7-point advantage for the 1B when neither model is allowed to reason — and a 20.8-point advantage for LFM3-VL-2B the moment it is allowed to. We-Math is a strict-accuracy benchmark on multi-step visual math, and the score does not move gradually between the two models. It flips. That is the clearest single illustration in the table of what the reasoning mode is buying.

The practical translation: if you disable reasoning for latency and your workload involves numbers in charts or diagrams, the larger model's aggregate advantage does not protect you. The 1B is the better tool.

Where LFM3-VL-2B wins regardless of mode

The reverse is also true, and the categories are just as concrete. Five of them are close to one-sided even in the no-thinking column.

• GUI understanding — 83.1 thinking, 81.5 no-thinking, vs 45.3 for MiniCPM-V 4.7 1B. ScreenSpot-Pro is 55.5 vs 4.9; ScreenSpot-v2 Desktop is 94.6 vs 51.5.

• Visual grounding — 77.3 thinking, 72.6 no-thinking, vs 51.1. RefCOCO-avg lands at 90.5 vs 86.3.

• Embodied reasoning — 64.4 thinking, 57.9 no-thinking, vs 45.3.

• Agentic — 42.5 thinking, 32.9 no-thinking, vs 16.9. The weakest category for all four models in the table, and the widest relative gap between these two.

• Multilingual — 84.4 thinking, 76.2 no-thinking, vs 77.4. The only category where the no-thinking Liquid model loses to the 1B.

ScreenSpot-Pro at 55.5 against 4.9 is the single largest gap between the two models anywhere in the 62 benchmarks. If you are building UI automation, there is no decision to make here; if you are building a maths tutor, the ranking reverses. A 13.9-point average is an artefact of averaging two populations that behave very differently.

Two harnesses, two numbers for the same benchmark

This is the part that makes cross-vendor comparison treacherous, and it deserves its own section because both vendors publish numbers for benchmarks with the same name.

OpenBMB reports MathVista 74.3 for MiniCPM-V 4.7 1B on its own page. Liquid's re-run of the same model on the same named benchmark reports 71.2. OpenBMB reports OCRBench 86.1; Liquid's harness reports OCRBench v1 at 82.1. Neither lab is lying. They are running different prompt formats, different answer-extraction rules or different subsets, and the resulting 3.1- and 4.0-point differences are ordinary harness variance.

The consequence is that you cannot take OpenBMB's 74.3 and subtract Liquid's 68.5 no-thinking MathVista figure for LFM3-VL-2B to get a 5.8-point win. That was a defensible reading before October 27, when the only shared benchmark between MiniCPM-V 4.7 1B and LFM2.5-VL-3B was measured on two different harnesses. It is not defensible now that Liquid has published both models inside one table. Use the table, and treat every vendor-page figure as belonging to its own instrument.

The tie that quietly settles an older argument

Inside that same table is a result nobody at either company has advertised. MiniCPM-V 4.7 1B averages 57.5 across the 62 benchmarks. LFM2.5-VL-3B averages 57.6.

OpenBMB's product page claims the 1B model surpasses LFM2.5-VL-3B on OCR, chart reasoning and video understanding. On Liquid's suite, that claim is half right and half wrong, and the aggregate is a dead tie:

• Document, OCR and chart — 73.0 for MiniCPM-V 4.7 1B vs 71.2 for LFM2.5-VL-3B. Claim supported, narrowly.

• Video — 53.7 vs 41.3. Claim supported, decisively.

• Math — 51.8 vs 40.6. Claim supported, and not mentioned in the original claim.

• Visual grounding — 51.1 vs 68.3. Claim fails badly.

• GUI understanding — 45.3 vs 66.0. Claim fails badly.

• Agentic — 16.9 vs 21.1. Claim fails.

So a 1.55B Apache 2.0 model matches a 3.1B model from a well-funded lab on a shared 62-benchmark suite, and does it by winning some categories decisively and losing others decisively. That is a real result about capability density, and it is also a warning about what an average hides. Which brings the comparison back to LFM3-VL-2B: the model that beats both is the one whose reasoning mode lifts its weakest categories, not the one that is uniformly better.

Licence, downloads, and what you can actually ship

The licence difference will decide more deployments than any row above. Apache 2.0 requires no conversation with a legal team. The LFM Open License v1.0 carries commercial conditions, and while it is open enough that Liquid ships every model in the family under it, it is not the same permission.

Adoption tells a similar story. On Hugging Face, MiniCPM-V 4.7 1B has accumulated roughly 155,000 downloads and just over 400 likes since October 6. LFM3-VL-2B, at the time of writing, sits at a few hundred downloads and 16 likes, having been public for nine days and announced for one. Download counts are a noisy signal — CI pipelines and quantisation mirrors inflate them — but a three-order-of-magnitude gap is not noise. It reflects that one model has a month of tooling, tutorials and community quantisations behind it and the other has a day.

Tooling follows the same pattern. MiniCPM-V 4.7 1B lists Transformers, SGLang, vLLM, llama.cpp, Ollama and OpenVINO, plus native iOS, Android and HarmonyOS applications. LFM3-VL-2B has SGLang at launch, GGUF weights available, and llama.cpp support still in review — with no video input in the pending implementation, per its own card.

Which model for which job

• UI automation, screen reading, robotics perception — LFM3-VL-2B, without hesitation, and the no-thinking mode is fine. ScreenSpot-Pro at 55.5 against 4.9 is not a margin you can engineer around.

• Chart reasoning, visual maths, diagram questions — MiniCPM-V 4.7 1B if you are running without reasoning, and a genuine contest if you are not. Check We-Math on your own data; a 27.7-point no-thinking gap is the kind of thing that shows up immediately in production.

• Long video — closer than the category averages imply. LFM3-VL-2B leads the Video category 61.2 to 53.7, but MiniCPM-V 4.7 1B actually wins Video-MMMU 50.1 to 49.6 even against the thinking configuration.

• Anything that has to clear legal without a review — MiniCPM-V 4.7 1B, on Apache 2.0 alone.

• Anything that needs a mature inference stack today — MiniCPM-V 4.7 1B, for the next few weeks at least, until llama.cpp merges LFM3-VL support.

Where a router fits when neither model is on it

Neither of these checkpoints is on OrcaRouter's catalogue; both are weights you download, quantise and run. The interesting use of a routing layer here is not serving them but composing around them. Model fusion lets a panel of models answer the same question together, and a local 1.55B specialist in visual maths paired with a hosted model that is stronger on the categories where it loses — grounding, agentic work — is a configuration that beats either model alone. The routing DSL is how you express which leg handles which question without rewriting the application each time a checkpoint ships.

The honest summary is that the 13.9-point headline is real, measured on one lab's harness, and misleading as a purchase decision. LFM3-VL-2B is the better model for screens and grounding by a margin nothing else in its size class approaches. MiniCPM-V 4.7 1B is the better model for visual mathematics without reasoning, ships under a licence that needs no negotiation, and ties a model twice its size on the same suite. The right answer depends on which of those sentences describes your workload — and neither vendor's average will tell you that.