MiniCPM-V 4.7 1B vs LFM2.5-VL 3B: OpenBMB Says Its 1B Beats a 3B. Here Is What That Claim Hides

MiniCPM-V 4.7 1B and LFM2.5-VL 3B are both open-weight vision-language models built for phones and small computers, and they sit less than three weeks apart on the calendar. MiniCPM-V 4.7 1B was uploaded by OpenBMB on October 6, 2026 and dated October 12 in the vendor's release notes. LFM2.5-VL 3B was published by Liquid AI on August 11, 2026. One is 1.55 billion parameters and Apache 2.0. The other is 3.12 billion parameters under the LFM Open License v1.0. And OpenBMB's own product page makes a claim that should stop anyone shopping on size alone: that its 1B model surpasses LFM2.5-VL 3B on OCR, chart reasoning and video understanding. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

That sentence is doing a lot of work. It is also, as far as anyone can check, not contradicted by the numbers — which is more interesting than if it were. This piece walks through what each vendor actually published, where the comparison is real, and the two places where the claim quietly changes the subject.

The claim, in the vendor's own words

OpenBMB's 1B page says the model "is a comprehensive upgrade over MiniCPM-V 4.6-1.3B and surpasses Qwen3.5-0.8B and LFM2.5-VL-3B, staying ahead of its size class in OCR, chart reasoning, and video understanding." The clause that matters is "staying ahead of its size class." The comparison to LFM2.5-VL 3B is offered as evidence of capability density; the sentence that follows limits the claim to a size class the LFM is not in.

Liquid AI's card does not return the favour. It does not mention MiniCPM at all, and it benchmarks its model against G​emma4 E2B and E4B, InternVL 3.5 2B and 4B, and Qwen3.5-2B and 4B. So this is a one-sided comparison: one vendor measured itself against the other, and the other has never acknowledged the matchup. That asymmetry is the first thing to know before reading any head-to-head, including this one.

What each vendor puts on the record

Liquid AI publishes a numeric table. LFM2.5-VL-3B reports MathVista 68.5, MMMU Pro 30.5, ChartQA 81.3 and OCRBenchv2 47.5 (English-only subset), alongside a sibling model's scores on the same rows. Those are vendor numbers, run by the vendor, and nobody independent has to agree with them — but they are falsifiable, which is what makes them useful. If you reproduce 55 on MathVista, you know something in your setup differs.

OpenBMB publishes a numeric summary in prose and full charts underneath it. For the 1B, the page states OCRBench 86.1, MathVista 74.3, CharXiv Reasoning 48.8, LongVideoBench 53.4 and MotionBench 54.55, and presents the rest of the comparison as images with a 4× and 16× column. The model card itself contains no benchmark table at all.

Compare the two on the one benchmark both name. MathVista: MiniCPM-V 4.7 1B at 74.3, LFM2.5-VL 3B at 68.5. The 1B is ahead by 5.8 points on the vendor's own reporting. On OCR, the two use different instruments — OCRBench versus OCRBenchv2's English subset — and the numbers are not comparable, which is exactly the kind of detail a "surpasses" sentence skips.

Where the comparison is apples to apples

• Parameters — 1.55B dense, BF16, 3.1 GB on disk vs 3.12B dense, BF16; the LFM is roughly twice the model

• Vision encoder — SigLIP2-style tower, 27 layers, 1152 hidden, 980-pixel tiles vs SigLIP2 NaFlex 400M, 512×512 non-overlapping patches plus a resized thumbnail

• Text backbone — a Qwen3.5-derived config with 24 layers, 1024 hidden, a 3:1 linear-to-full attention pattern vs LFM2.5-2.6B, a hybrid architecture of Liquid AI's own design

• Context — 262,144 tokens vs 32,768 tokens

• Visual token budget — switchable 4× or 16× compression per request vs a fixed NaFlex patch budget set by resolution

• Licence — Apache 2.0 vs LFM Open License v1.0, which carries commercial terms rather than plain permissiveness

• Languages named — English and Chinese in the card's examples vs 16 languages listed explicitly (Arabic through Vietnamese)

• Published evaluation — five prose figures plus chart-only comparisons vs a numeric table across four benchmarks and six sibling models

• Independent evaluation — no Artificial Analysis entry for either as of this writing vs no Artificial Analysis entry either

• Vendor-reported speed — no tokens-per-second figure for the 1B on any device vs 228 tok/s on an Apple M5 Max, 116 tok/s on an AMD Ryzen AI Max+ 395, 20 tok/s on a Galaxy S26 Ultra

• Memory — about 3.1 GB BF16; Q4 GGUF 0.67 GB plus a 0.90–1.11 GB vision projector vs under 3.3 GB as published, with GGUF, ONNX and MLX builds offered

• Serving paths named — Transformers, SGLang, vLLM, llama.cpp, Ollama, OpenVINO, DSpark speculative decoding vs Transformers, vLLM, SGLang, llama.cpp, plus GGUF, ONNX and MLX exports

• On-device story — iOS, Android, HarmonyOS apps and day-0 Rockchip, Houmo, Intel, AMD and Arm support, demonstrated on an iPhone 17 Pro Max vs a Galaxy S26 Ultra decode figure and an MLX build for Apple Silicon

Two rows deserve a second look. The context window is an eightfold difference in MiniCPM's favour, and it is the kind of number that looks decisive until you remember that a 1B model holding 262k tokens is holding a KV cache larger than its own weights. In practice, on a phone, neither model will run anywhere near its ceiling. And the licence row is the one that decides commercial adoption more often than any benchmark: Apache 2.0 needs no conversation with a lawyer, and the LFM Open License does.

The part the claim changes the subject on

OpenBMB is comparing its model to a competitor twice its size, on benchmarks it chose, and reporting the result in prose. That is not dishonest — the numbers are published and the comparison is stated. But it is answering a question nobody deploying a model actually asks. Nobody chooses between a 1B and a 3B because they want the smaller one to be better. They choose because one fits their device and the other does not, or because one has the language coverage they need, or because one's licence clears legal review.

On the first of those, the honest reading is that both fit. LFM2.5-VL-3B's card claims under 3.3 GB and demonstrates 20 tokens per second on a Galaxy S26 Ultra. MiniCPM-V 4.7 1B at Q4 with its projector lands near 1.8 GB. So the 1B's real advantage is not "it runs on a phone and the other does not" — it is headroom: roughly a gigabyte and a half less memory, which on a mid-range Android device is the difference between an app that survives and one the system kills.

On the second, the comparison inverts. LFM2.5-VL-3B names 16 languages in its card and ships ONNX and MLX builds alongside GGUF. MiniCPM-V 4.7 1B's documentation is oriented around English and Chinese examples, and its on-device story is oriented around iOS, Android and HarmonyOS chip support rather than around language coverage. If your product ships in Turkish, Polish or Vietnamese, the bigger model with the more restrictive licence is the one with the published claim to support you.

On the third, Apache 2.0 wins outright and there is not much to discuss.

What each is genuinely better at

LFM2.5-VL-3B is built around a specific set of jobs, and its card says so: single-turn, high-throughput, low-latency work such as near-real-time object detection in automotive settings, batch OCR of scanned documents with layout annotation, and on-device translation of menus and road signs. The card also names what the model is not for — long-context and reasoning-intensive tasks, visual web design, technical questions about blueprints. That kind of negative guidance is rare and worth trusting more than the positive claims.

MiniCPM-V 4.7 1B is built around capability density at the bottom of the size range. The mixed 4×/16× token compression is the clearest expression of it: the same checkpoint can serve a detailed document read and a cheap thumbnail pass, chosen per request. The slicing controls (36 slices for images, 1 for video), the frame-sampling parameters for video, and the separate PDF-parsing and grounding cookbook entries all point at a document-and-screen workload rather than a detection workload. It also ships DSpark speculative-decoding heads as a separate 213M-parameter artefact, which is a real throughput lever if your runtime supports it and dead weight if it does not.

The practical split: if you are detecting objects in a camera feed or translating signage at 20 frames a second, the LFM's latency figures and language coverage are the argument. If you are turning screenshots, receipts and page images into structured text with a token budget you control, the MiniCPM's compression switch is the argument.

The gap a buyer should price in

LFM2.5-VL-3B has been out since August 11, 2026 — ten weeks at the time of writing — which in open-weight terms is long enough for the quantisation ecosystem to settle. GGUF, ONNX and MLX builds ship from the vendor itself. Any runtime you are likely to use has a working path.

MiniCPM-V 4.7 1B is three weeks old, and its card opens with a caveat about CUDA version mismatches breaking video decoding through torchcodec, with two documented workarounds. The architecture is new — a model_type that did not exist a month ago — and the minimum Transformers version is 5.18.0. None of that is a defect. It is the standard tax on adopting a model in its first month, and the reason a quiet release with chart-based evaluation is a good candidate for a side-by-side test and a questionable candidate for a production path in week one.

The cheap way to run that test is to keep the incumbent serving while you evaluate the newcomer, which is a routing problem rather than a modelling one. OrcaRouter runs more than 200 hosted models behind one OpenAI-compatible endpoint at 0% markup, passing each provider's list price through unchanged, so a price cut from a vendor shows up in your bill the same day rather than after a repricing cycle. Its routing DSL can hold a primary and a fallback in a single call, and automatic failover keeps the path alive when a provider has a bad hour — which is what you want when the model on trial is a 1B local checkpoint and the model in production is something you are paying for. Neither model in this article is in that catalogue: both are checkpoints you run yourself, and the router sits around them rather than offering a way to call them.

The claim, settled

Does MiniCPM-V 4.7 1B beat LFM2.5-VL 3B? On MathVista, by 5.8 points as both vendors report it. On OCR, the two published different instruments and the comparison cannot be made. On chart reasoning and video understanding, OpenBMB asserts a win and publishes it as a chart, and Liquid AI has never responded. The one comparison that is fully checkable supports the claim; the others are the vendor grading its own homework in a format that resists quotation.

What the claim does not survive is being read as "the smaller model is simply better." LFM2.5-VL-3B offers sixteen languages, exported ONNX and MLX builds, a documented speed figure on a shipping Android phone, and a card that tells you where not to use it. MiniCPM-V 4.7 1B offers a switchable token budget, a 262k context window, Apache 2.0, and about 1.8 GB of quantised memory. Those are different products dressed in the same benchmark row, and the right answer depends on which of those lists matches your deployment — not on which number is larger.

Test both on your own images at your own quantisation. The weights are free and the afternoon is the only cost.