Molmo2-4B vs MiniCPM-V 4.7: Neither Model Sits Where You Think It Does

The matchup in the title does not exist in the weights. Molmo2-4B is a single 4.85B checkpoint. MiniCPM-V 4.7 is not one model but two, and neither is a 4B: a 1.55B dense variant built for phones and a 35.2B mixture-of-experts variant with roughly 3B active parameters and a 262,144-token context. So a reader searching for "Molmo2-4B vs MiniCPM-V 4.7" is really asking a three-way question — is the 4B slot worth occupying at all, or do you go down to 1B for the footprint or up to the MoE for the capability? That is a more useful question than the one that was typed, and it has a clearer answer. A three-model decision is a routing problem before it is a modelling one. The routing DSL composes several models into one call instead of one contract per model.

The dates matter here too, because both models are frequently described as newer than they are. Molmo2-4B was announced by the Allen Institute for AI on 11 December 2025, with weights on Hugging Face from 14 December. MiniCPM-V 4.7's weights went up on 6 and 7 October 2026, the model card followed on 12 October and an English press release from ModelBest and OpenBMB on 15 October. The recent change on the Molmo side is not a new model: Molmo2 became a native architecture in upstream Transformers when PR #43451 merged on 9 October 2026 and shipped in v5.20.0 on 15 October, and on 5 November Ai2 published a Molmo2-4B fine-tune for a robotics testbed. Read either of those as a reason to revisit the model. Neither is a launch.

Three checkpoints, three positions on the curve

• Parameters — Molmo2-4B is 4.85B; MiniCPM-V 4.7-1B is 1.55B; MiniCPM-V 4.7-35B-A3B is 35.2B total with about 3B active per token.

• Architecture — Molmo2-4B is a SigLIP 2 tower bolted to Qwen3-4B-Instruct-2507; the MiniCPM-V 4.7 pair use intra-ViT early compression from LLaVA-UHD v4 with mixed 4x/16x visual token compression, and the 35B variant is the first MoE in the MiniCPM-V line.

• Context — 36,864 positions for Molmo2-4B; 262,144 tokens for both MiniCPM-V 4.7 variants.

• Modalities — Molmo2-4B handles image, multi-image and video, and emits pointing coordinates in 0–1000 image space; MiniCPM-V 4.7 handles single image, multi-image and video, and answers in text.

• Licence — Molmo2-4B is Apache 2.0 on the weights, with the card warning that its training data is subject to academic and non-commercial research use only; both MiniCPM-V 4.7 variants are Apache 2.0.

• Distribution — Molmo2-4B is float32 safetensors, 19.42 GB, four shards, no official quantizations; MiniCPM-V 4.7 ships GGUF, int4, AWQ, GPTQ, FP8 and DSpark variants across both sizes.

The distribution row is the one that decides real projects. A 262,144-token context is only useful if you can afford to hold it, and OpenBMB has done the unglamorous work of publishing the quantized formats people actually deploy — the official 1B GGUF at Q4_K_M is 672 MB, which is the difference between a demo and something that runs on a device you already own.

Where the 4B actually wins

Molmo2-4B's case rests on two things the MiniCPM-V 4.7 pair do not do the same way: pointing as a first-class output, and video tracking with timestamps.

The pointing format is a numeric point in 0–1000 coordinates, wrapped in a tag the model card ships a regex for. Multi-image pointing flattens to a list of frame-index, x, y triples. That is a different contract from "describe where the object is," and it changes what you can build: crop pipelines, region-of-interest follow-ups, count-then-verify loops, and any workflow where a downstream stage needs a coordinate rather than a sentence.

Video is where the parameter budget goes. Ai2's card names video captioning QA, video subtitle QA, video pointing, video tracking and two multi-image pointing sets among its training data, and the released video preprocessor defaults to 384 frames at 2 fps at 378×378 with 3×3 pooling. MiniCPM-V 4.7 does support video — OpenBMB reports LongVideoBench 68.9 for the 35B-A3B and 53.4 for the 1B — but it is a text-out model on that path, and a long video competes with the prompt for the same token budget.

Where MiniCPM-V 4.7 wins, twice

At the bottom of the range, the 1B is a different class of object. At 1.55B parameters it is roughly a third of Molmo2-4B's size, it runs under llama.cpp and Ollama, and OpenBMB reports OCRBench 86.1, MathVista 74.3 and CharXiv 48.8 for it. Those are vendor numbers on a vendor-selected harness, and they should be read that way — but a 1.55B model that does full-page OCR credibly is a genuinely different procurement decision from a 4.85B model that needs a GPU.

At the top of the range, the 35B-A3B makes the 4B question look parochial. OpenBMB's published figures for it include MMMU 80.2, OCRBench 91.2 at 4x visual tokens and 90.40 at 16x, CharXiv 75.6, LongVideoBench 68.9, MathVerse 86.5 and MathVista 88. It also claims something more interesting than a benchmark: with Lookahead Expert Scheduling, layer-wise quantization and DSpark speculative sampling, the 35B-A3B decodes at 13+ tokens/s in about 3 GB of memory on an iPhone 17 Pro Max. Treat that as a vendor claim about a specific toolchain rather than a general property of the model — but if it holds, it removes the usual reason to prefer a dense 4B on device.

All of these are vendor-reported. Neither OpenBMB nor Ai2 has published results for its model on the other's benchmark suite, so no row above is a head-to-head.

The deployment menu is the underrated difference

MiniCPM-V 4.7 lists SGLang, vLLM, llama.cpp and Ollama for inference, SWIFT and LLaMA-Factory for fine-tuning, and day-0 support for Rockchip, Houmo, Intel, AMD and Arm. Molmo2-4B's card lists Transformers with remote code, a separate molmo_utils dependency and decord2 for video, and AI2's own platform docs point at vLLM nightly rather than stable. The gap narrowed in October, when native Transformers support landed in v5.20.0 and the remote-code requirement became optional, but it is still a gap: one project has a quantization matrix, the other has a research release.

This is where routing earns its place rather than being a slogan. A team that wants the MiniCPM-V 4.7 1B on device for the cheap path and a larger model for the hard path does not need two contracts to do it — the routing DSL composes several models behind one endpoint, and 0% markup means the provider's list price is what you pay, so a vendor price cut is live on the same day. That matters more for a three-model decision like this one than for a straight A-or-B.

How to actually choose

If the constraint is memory, take MiniCPM-V 4.7-1B. Nothing in the Molmo2-4B package gets near 672 MB, and no amount of pointing accuracy compensates for a model that does not fit.

If the constraint is capability per dollar and you can serve a MoE, take MiniCPM-V 4.7-35B-A3B. On OpenBMB's own numbers it is in a different league from either smaller model, and the active-parameter count keeps the per-token cost closer to a 3B than a 35B.

If the requirement is a coordinate, take Molmo2-4B. Points with timestamps, multi-image pointing and tracking are the reasons the model exists; nothing in the MiniCPM-V 4.7 documentation describes an equivalent output contract.

And if you genuinely need a 4B-class generalist that does image, multi-image and video without a 19 GB float32 download, the honest answer is that this particular matchup does not offer one — Molmo2-4B is the only model in the comparison in that slot, and it enters with the thinnest deployment tooling of the three. What would change that: an official quantized Molmo2-4B, or a 4B-class MiniCPM-V. Neither exists today.