Molmo2-4B vs Moondream 3: One Hosted Option Closes on 13 November

There is a date attached to this comparison, and it is one week away. Moondream Cloud — the inference and Lens side of the business — shuts down on 13 November 2026, thirty days after Moondream announced it had been acquired by Prime Intellect. Unspent purchased credits are being refunded. If you built on the Moondream 3 hosted endpoint, the practical question stopped being "Molmo2-4B or Moondream 3" and became "which set of open weights do I run myself." Both models survive the shutdown. Only one of them ever had a hosted API to lose.

Worth being precise about which Moondream is being compared, because the name is used loosely. Moondream 3 Preview went up on Hugging Face in September 2025. The current release is Moondream 3.1, published on 7 July 2026 as moondream3.1-9B-A2B, and that is the checkpoint this article treats as "Moondream 3." Molmo2-4B is the other side: announced by the Allen Institute for AI on 11 December 2025, weights live on 14 December, Apache 2.0 on the weights with an explicit warning that the training data is subject to academic and non-commercial research use only. It has never had a hosted API from Ai2, and that turns out to matter.

What is genuinely recent on the Molmo side is not a release. On 9 October 2026 the Molmo2 architecture was merged into upstream Transformers and it shipped in v5.20.0 on 15 October, so the model no longer requires trust_remote_code=True. On 5 November Ai2 published Molmo2-4B-tb, a fine-tune of this checkpoint for a block-stacking robot. Neither makes Molmo2-4B new, and neither changes its footprint.

What survives, and what does not

The weights are not going anywhere. Moondream ships moondream3.1-9B-A2B under the Moondream Model License 1.0 and publishes Photon, its own inference engine, for local deployment on Ampere-or-newer NVIDIA GPUs and Apple Silicon. The licence permits commercial use, fine-tuning, quantization, redistribution, edge deployment and customer-controlled VPCs. What it forbids is one specific thing: providing a general-purpose hosted Moondream service to third parties without a separate commercial licence.

That clause is the reason the shutdown matters more than it first appears. A team that was calling the Moondream Cloud API cannot simply stand up the same endpoint themselves for their customers. They can run Moondream inside their own product, and they can run it on infrastructure they control, but "Moondream as a service" is exactly what the licence reserves. If the endpoint was the product, 13 November is a licensing event and not just a migration.

Molmo2-4B has no equivalent trap, because there was never a hosted version to lose. The weight licence is Apache 2.0, and the restriction lives in the data statement instead: Ai2's card says the model "is trained on third party datasets that are subject to academic and non-commercial research use only," and directs you to review the sources. Neither model is a clean "just ship it" — one restricts what you can offer, the other restricts what the training lineage supports.

Two different machines under the hood

• Parameters — Molmo2-4B is 4,850,869,200 dense; Moondream 3.1 is 9.0B total with 2.0B active per token, a mixture of eight experts with top-2 routing plus one shared expert.

• Weight size — 19.42 GB of float32 safetensors for Molmo2-4B, four shards; Moondream 3.1 is distributed for Photon and via Hugging Face, with the vendor emphasising a small runtime footprint rather than a large download.

• Context — 36,864 positions for Molmo2-4B; 32,768 tokens for Moondream 3.1.

• Input surface — Molmo2-4B takes image, multi-image and video; Moondream 3.1 takes one image per call and has no native video path.

• Output contract — Molmo2-4B returns points scaled to 0–1000 of image dimensions, with timestamps in video mode and a documented regex for parsing; Moondream 3.1 exposes four named skills — query, detect, point, caption — each returning structured JSON.

• Licence — Apache 2.0 weights with a third-party data caveat for Molmo2-4B; Moondream Model License 1.0 for Moondream 3.1.

The output row is the most interesting difference and the least discussed. Molmo2-4B's points are a raw coordinate stream you parse and post-process. Moondream 3.1's skills are an API contract: detect returns objects, point returns points, caption returns a caption, and the shape does not vary with the prompt. One is a model you build on; the other is a service you call. That design difference is a better predictor of which one fits your codebase than any benchmark row.

Scores that share no harness

Moondream publishes a substantial table for 3.1, and it is worth quoting precisely because most summaries round it. Vendor-reported, all on Moondream's own evaluation: COCO detection [email protected] 81.46, ODinW-13 93.94, LVIS 67.4, Dense200 74.61, SKU-110K 52.77, CrowdHuman 74.5, RefCOCO-M 87.1, HumanRef 70.39, DOTA-v2 59.63, CountBench 90.35, TallyQA 74.86, PixMo-Count 88.68, ChartQA 86.01 and DocVQA 88.6, for a stated average of 77.9. The same post reports 34.2 requests per second and first-token latency of roughly 20–30 ms on Cloudflare Workers AI, with a point call around 145 ms end to end.

Molmo2-4B reports one aggregate: 62.8 average across 15 academic benchmarks, in a table that places it behind the 8B sibling at 63.1 and alongside Qwen3-VL-4B at 58.1, InternVL3.5-4B at 53.4 and Eagle2.5-8B at 60.7, while the closed models in the same table sit at 70.0 for Gemini 3 Pro and 70.6 for GPT-5.

Those two averages cannot be compared. They cover different tasks, Moondream's is heavily weighted toward detection and counting where it is strongest, and neither vendor has run its model on the other's suite. Anyone presenting 77.9 against 62.8 as a six-point win is comparing two different exams and calling it a score.

The count that both vendors care about

Counting is the one place where the two models' claims genuinely overlap, and where the honest reading is uncomfortable for both. Moondream reports CountBench 90.35 and PixMo-Count 88.68. Ai2's announcement says the Molmo 2 family leads open-weight models on video counting, and then concedes in the same breath that no model in its evaluation reaches 40% accuracy. Those statements are not in conflict — one is dense-object counting on curated images, the other is counting across video — but they mean "good at counting" is not a portable claim. Specify the test before you trust the number.

If you are deciding this week

Start from where the request will run. If you were calling the Moondream Cloud API, the migration is now a self-hosting project, and the licence decides whether you can re-expose it. Run Photon locally, keep Moondream inside your product, and do not rebuild the endpoint for third parties without a commercial licence.

If video is in scope, Moondream 3.1 is out. It takes one image per call, and no amount of skill structure gets around that. Molmo2-4B's 384-frame default at 2 fps is the reason to accept its 19.42 GB float32 download and cast it down before serving.

If your pipeline wants coordinates but not a paragraph, both can do it, and the choice becomes packaging: Molmo2-4B's tagged point stream with a regex you own, or Moondream 3.1's point skill returning JSON. The second is less work today and carries a licence obligation tomorrow.

For teams that were using Moondream as one of several hosted vision models rather than as the product, the migration is less dramatic than it sounds: OrcaRouter passes the provider's own list price with no markup added across more than 200 model ids, with automatic failover when an upstream endpoint degrades, so replacing one hosted vision call with another does not mean a second contract or a rewritten client. Neither Moondream 3.1 nor Molmo2-4B is in that catalogue — both are self-hosted weights — but the rest of the stack usually is.

The clearest thing to say about this pairing is that it was never a fair fight and is now a smaller one. Moondream 3.1 is a fast, well-packaged vision API with a licence that reserves the service business for its owner. Molmo2-4B is a research release with a video-native point output and no packaging at all. The shutdown does not make one better than the other. It just removes the option of not choosing.