WeVisOne-VL-2B vs MiniCPM-V 4.7 1B: 0.45B Apart, Different Jobs

The names suggest a two-to-one gap. The checkpoints do not. MiniCPM-V 4.7 1B ships 1,554,476,272 parameters in bfloat16, and WeVisOne-VL-2B ships about 2 billion — so the real distance between them is roughly 0.45 billion parameters, not a billion. Put them on the same disk and they would take up comparable space. Put them on the same task and they fall apart completely, because MiniCPM-V 4.7 1B was built to read a document, watch a video and hold a conversation about it, while WeVisOne-VL-2B was built to find an object and hand back a mask.

Both are Apache 2.0. Both are small enough to interest someone building on a single GPU. Both are recent — MiniCPM-V 4.7 1B uploaded on 6 October 2026 and dated 12 October in OpenBMB's release notes, WeVisOne-VL-2B on 22 October. That combination of similarity and difference is what makes the pairing worth a page: almost every headline spec is close, and almost every actual use is not.

First, the names, corrected

OpenBMB's naming is loose in a way that has already caused confusion in comparison tables. The model called 1B has 1.55 billion parameters by its own safetensors metadata; vendor materials have quoted 1.3B and 1.38B in different places. The 262,144-token context is real and is set in the model config. The architecture is a 24-layer, 1024-hidden text backbone derived from Qwen3.5 with a 3:1 pattern of linear to full attention, paired with a 27-layer SigLIP2-style vision tower working on 980-pixel tiles, plus a selectable 4× or 16× visual-token compression mode that trades detail for cost per request.

Tencent's naming is tighter. WeVisOne-VL-2B is a 2-billion-parameter model built on Qwen3-VL, with a separate 9,511,460-parameter pixel decoder that the card calls "a required core component, not an optional add-on." The decoder is not a rounding error in the deployment story: without it the model produces boxes and points, and none of the mask or matte outputs exist.

So the correct framing is not small versus large. It is 1.55B versus 2B — a 29 percent difference in weights — with entirely different output contracts.

Two benchmark vocabularies that never meet

This is the part most comparisons get wrong, so it is worth stating before any number appears. OpenBMB evaluates MiniCPM-V 4.7 1B on general multimodal benchmarks: OCRBench 86.1, MMMU 56.89, MathVista 74.30, MMBench 83.17, MMStar 69.53, MMMU-Pro 38.03, MathVerse 60.41, CharXiv Reasoning 48.8, LongVideoBench 53.4 and MotionBench 54.55. Those are vendor-run figures, presented mostly as charts rather than as a table with harness details, and no independent body has reproduced them.

Tencent evaluates WeVisOne-VL-2B on perception benchmarks and reports no general multimodal scores at all. There is no MMMU, no MMBench, no MathVista, no DocVQA and no captioning result anywhere in its seven tables. What it reports instead is task-specific: detection F1, counting exact-match, pointing F1, OCR localization F1, segmentation cIoU and matting error.

The collision is sharpest on OCR, because both models claim to do it. MiniCPM's 86.1 is OCRBench — a transcription score, asking whether the model read the characters correctly. WeVisOne's OCR numbers are localization F1 at a 0.5 IoU threshold: 62.1 on SROIE, 42.6 on ICDAR2015, 42.5 on TotalText and 30.5 on HierText, measuring whether the model drew the right box around the text. A model can be excellent at one and mediocre at the other. Printing "86.1 versus 62.1" would be comparing a spelling test to a pointing test.

There is no benchmark on which both models report a number computed the same way. That is not a rhetorical flourish — it is the central finding of this comparison, and it means any ranking you see of these two is assembled from incommensurable parts.

What 262,144 tokens buys, and what WeVisOne won't say

MiniCPM-V 4.7 1B publishes a 262,144-token context and a video path that accepts up to 128 frames by default, with one slice per frame. That is a document-and-video machine: long PDFs, multi-page scans, a minute of footage. The 4× and 16× compression switch exists precisely because holding that much visual detail is expensive, and OpenBMB lets you choose how much of it to keep per request.

WeVisOne-VL-2B publishes no context length. The card is silent; the paper's training stages ran at 4,096 and 8,192 tokens, and inference caps each image at 1,960 visual tokens with greedy decoding. None of those is a deployment window, and Tencent does not pretend one is. It also rejects video grids outright — the pixel decoder is image-only, single-turn, no chain-of-thought.

That asymmetry tells you what each vendor optimised for. OpenBMB optimised for how much input one model can swallow. Tencent optimised for how precisely one model can point at what it swallowed, and then drew a hard boundary around the input type to keep the decoder simple.

Where each one actually runs

MiniCPM-V 4.7 1B has the wider deployment surface by a distance: Transformers 5.18 and later, vLLM, SGLang, llama.cpp and Ollama, with GGUF, AWQ, GPTQ, FP8, bitsandbytes and OpenVINO builds, plus iOS, Android and HarmonyOS support and day-zero chip enablement from a list of vendors that includes Rockchip, Arm and Apple. OpenBMB also ships a separate 213-million-parameter DSpark draft model for speculative decoding.

WeVisOne-VL-2B has its own GitHub package, a Transformers path, and vLLM and SGLang listed on the card. No GGUF. No quantized build. No ONNX. No mobile story. And the dense output is a stride-1, full-resolution alpha matte, which means memory scales with the image rather than with the parameter count — a 2B model that is cheap to hold and not necessarily cheap to run at 4K.

If your target is a phone, this section ends the comparison. MiniCPM was built for that constraint and says so in its formats list. WeVisOne was not, and does not claim to be.

The one place they tie: the licence

Apache 2.0 on both. No revenue threshold, no field-of-use restriction, no separate commercial agreement. In a category where Liquid AI uses its own revenue-capped licence and Moondream uses a source-available licence with a no-third-party-service grant, two Apache 2.0 vision models at this size is worth noticing — it means the licence is not the deciding variable here, which is unusual enough to say out loud.

The tie also removes the most common reason teams pick the smaller model. If both clear legal review, you are choosing on capability and deployment, not on terms.

Choosing between them

Ask what the output feeds. If it feeds a document store, a chat window, a video index or a downstream language model, MiniCPM-V 4.7 1B is the shape you want: it produces text at length, it holds a long input, and it ships in formats that run on consumer hardware.

If it feeds a compositing pipeline, an image editor, a catalogue cleanup job or a region-of-interest crop, WeVisOne-VL-2B is the shape you want: boxes, points, binary masks and continuous alpha mattes, plus a rejection behaviour that matters more than any benchmark. Tencent reports a 50.2 rejection rate on RefCOCOg for absent-object queries, against 0.5 and 0.6 for two specialised grounding models on the same table. If most of your images do not contain the thing you are looking for, that number is the one that saves you money.

There is a third case where they compose rather than compete: a document pipeline where MiniCPM reads the page and WeVisOne finds the region — the signature, the stamp, the product photo — that the reading step needs to isolate. That is not a routing problem so much as a pipeline-design one, and it is where the OrcaRouter product is genuinely useful rather than bolted on. Neither of these models is on our catalogue; both are self-hosted, and we are not going to imply otherwise. What we provide is the hosted half: more than 200 models behind one API key at provider list price with no markup, automatic failover when a provider degrades, and model fusion, which lets a panel of models answer the same question together — useful when a transcription is about to enter a database and you want a second opinion before it does.

The number to remember

Not 1B and 2B — 1.55B and 2B. The naming makes them look like different weight classes and the checkpoints make them look like near-neighbours, and both impressions are misleading. What separates MiniCPM-V 4.7 1B from WeVisOne-VL-2B is not size, licence or even recency. It is that one returns words and the other returns pixels, and no benchmark either vendor published can be subtracted from the other to tell you which is better at being the thing you need.