MiniCPM-V 4.7 vs Moondream 3.1 9B A2B: A Week Before One of Them Loses Its API

On 13 November 2026, Moondream Cloud shuts down. That is thirty days after Moondream announced it was joining Prime Intellect, and it is the single most decision-relevant fact in any comparison between Moondream 3.1 9B A2B and MiniCPM-V 4.7 today. One of these models is a hosted service that is about to stop being hosted. The other was never a hosted service to begin with: OpenBMB published MiniCPM-V 4.7 as open weights on 6 and 7 October 2026, added the model card and an Apache-2.0 licence on 12 October, and left serving to you.

That asymmetry inverts the usual comparison. On a spec sheet, Moondream 3.1 9B A2B is a 9-billion-parameter sparse model with about 2 billion active, built around structured perception primitives, and MiniCPM-V 4.7 35B-A3B is a 35.2-billion-parameter mixture of experts with roughly 3 billion active, built around reading images, documents and video in a 262,144-token context. On a procurement sheet, one has a published per-token price for another seven days and the other has no price at all. Both facts are true at once, and which one you weight depends on whether you are prototyping or operating.

What each model is actually for

Moondream 3.1 9B A2B shipped on 7 July 2026. According to the vendor's own model page it is a sparse mixture of experts — eight experts with top-2 routing plus one shared expert, 9.0B total and 2.0B active per token — with a 32,768-token context and a deliberate restriction: one image per call, text out, no native video and no multi-image reasoning. What it does instead of prose is answer in structure. Its native skills are query, caption, point, detect and segment, and the point and detect outputs are coordinates, which is why the benchmark table is full of detection F1 rather than caption scores.

MiniCPM-V 4.7 is the opposite bet. It takes single images, multiple images and video, holds a quarter of a million tokens of context, and answers in language. OpenBMB's published numbers are comprehension numbers: 80.2 on MMMU, 91.2 on OCRBench, 75.6 on CharXiv reasoning, 68.9 on LongVideoBench, 86.5 on MathVerse. There is no point, detect or segment primitive in the model card. If you want a box around the cat, you ask for it in words and parse the words.

• Context — 32,768 tokens on Moondream 3.1 9B A2B against 262,144 on MiniCPM-V 4.7 35B-A3B.

• Inputs — one image per call for Moondream, against single-image, multi-image and video for MiniCPM.

• Output shape — structured query, caption, point, detect and segment results for Moondream, against free text for MiniCPM.

• Licence — the Moondream Model License 1.0 permits commercial use, self-hosting and fine-tuning, and restricts offering general-purpose Moondream inference as a hosted service; MiniCPM-V 4.7 is Apache-2.0 with no such clause.

• Weights on disk — a 9B sparse checkpoint for Moondream, against about 70 GB of BF16 weights across 16 shards for MiniCPM-V 4.7 35B-A3B, or 3.1 GB for the 1B sibling.

The 13 November cliff, and what actually changes

The acquisition was announced by Prime Intellect on 13 October 2026 and by Moondream a day later. The open weights stay open under the same licence, self-hosting through Photon stays free, and Moondream continues as a brand for small multimodal models inside Prime. What closes is the managed cloud: inference and Lens fine-tuning both end on 13 November, with unspent credits refunded and customers pointed at the acquirer's platform.

For a team that built against the hosted endpoint, that is a migration with a date on it. The weights do not disappear, but the operational shape does: no per-token bill, no managed scaling, no vendor-side latency guarantee. The published cloud price list puts Moondream 3.1 9B A2B at 0.30 dollars per million input tokens and 1.00 dollar per million output tokens on the real-time tier, half that in batch, which was cheap enough that self-hosting rarely pencilled out for low volume. After 13 November the comparison is not price against price; it is a GPU you own against a model you call.

MiniCPM-V 4.7 never had that managed layer. The only first-party serving path is OpenBMB's own ModelBest API, which is what the esp32-eye cat-monitor build uses for its 35B identity checks, and the rest is vLLM, SGLang, llama.cpp or Ollama on hardware you provide. That is more work up front and less exposure to someone else's corporate calendar. Given the last thirty days, the second property is worth something.

Benchmarks that never meet

There is no shared benchmark between these two models, and the two tables measure different competences. Moondream's numbers are its own, measured with its own Photon engine on a single H100 at batch 16, and the tasks are detection and counting: 81.46 on COCO, 93.94 on ODinW-13, 74.61 on Dense200, 90.35 on CountBench, 88.68 on PixMo-Count, 87.1 on RefCOCO-M. Its document numbers — 88.6 DocVQA, 86.01 ChartQA — are respectable and well below its detection scores. Its maths is simply absent: the release notes report no GSM8K or other arithmetic benchmark at all, so there is no maths number to compare and no evidence either way.

MiniCPM's numbers are OpenBMB's, published as prose bullets and chart images on the vendor's model page, with no numeric table and no independent reproduction. Its strength is the inverse of Moondream's: 91.2 OCRBench, 75.6 CharXiv reasoning, 68.9 LongVideoBench, 80.2 MMMU. It publishes no detection F1, no counting score and no matting or segmentation result, because it does not do those things natively.

The one place they can be compared is OCR, and even there the harnesses differ. Moondream reports 88.6 on DocVQA, a document question-answering score; MiniCPM reports 91.2 on OCRBench, a text-recognition score. Those are not the same test and should not be subtracted. Anyone who puts them in one column is producing a number that means nothing.

Where each one breaks

Moondream's documented limits are specific and useful: text smaller than about ten pixels, objects smaller than about twelve pixels, more than fifty instances in a frame, and no native video. Those are the boundaries of a perception primitive, and they are honest. If your images are product photos, shelf shots, aerial frames or crowded scenes, they describe your failure modes before you hit them.

MiniCPM-V 4.7's limits are less crisply documented, which is itself a finding. The model card claims on-device inference at 13 or more tokens per second on an iPhone 17 Pro Max using about 3 GB of memory, a figure that depends on Lookahead Expert Scheduling, layer-wise quantisation and DSpark speculative decoding all working together, and which no third party has reproduced. The esp32-eye build, the most detailed public deployment, found the 1B sibling unreliable at distinguishing four similar cats and moved identity to the 35B. That is a useful negative result: the small variant is a presence detector, not a fine-grained recogniser.

There is also a hardware asymmetry. Moondream 3.1 is a 9B sparse model that runs comfortably on a single modern GPU through Photon, and its engine is the reason the vendor can quote 34 requests per second at batch 16. MiniCPM-V 4.7 35B-A3B wants either a lot of VRAM or the streaming-expert machinery OpenBMB built for phones, and the 1B sibling wants almost nothing. Choosing between them on hardware is really choosing between one mid-size model and a two-size family.

If you are implementing this week

The decision tree is short. If your task is "find, count or outline objects and return coordinates", Moondream 3.1 9B A2B is the more direct tool, and the 13 November date only changes where you run it: self-hosted through Photon, under a licence that forbids reselling general-purpose Moondream inference. If your task is "read this document, this dashboard or this video and tell me what it says", MiniCPM-V 4.7 35B-A3B is the better fit, and it was always a self-hosting decision.

If you want neither commitment, the pragmatic path is a hosted general-purpose model for the reading work while you evaluate the open weights on your own data. That is where a router earns its place: OrcaRouter exposes more than 200 models behind one key at 0% markup on provider list prices, with automatic failover, so a pipeline is not tied to a single upstream's continued existence. Neither Moondream 3.1 9B A2B nor MiniCPM-V 4.7 is in that catalogue today, so this is advice about the surrounding architecture, not a claim that we serve either model.

What the last month has shown is that model availability is now a supply-chain question. A model can be excellent, open-weight and well documented, and still lose its managed endpoint on thirty days' notice because the company that built it was acquired. The teams that handled that well were the ones whose code had a base URL in a config file and a second option behind it.

The fair summary is that these are not substitutes. Moondream 3.1 9B A2B is a perception component with a clock running on its hosted form; MiniCPM-V 4.7 is a reading model with two sizes and no hosted form to lose. Pick the job first, then pick the model, then check whether the thing you picked will still answer next Friday.