Moondream 3 Preview's Hosted Endpoint Shuts Down on November 13

M​oondream 3 Preview was released on 18 September 2025, and the endpoint that serves it stops answering on 13 November 2026. Those are the two dates that matter in this story, and only the second one is urgent. The model itself is not being withdrawn — the weights stay on Hugging Face and the licence does not change — but the hosted path that most teams actually call, shared by M​oondream 3 Preview and its successor M​oondream 3.1, is being turned off, and the notice said so in a single line: on November 13, 2026, 30 days from today, the M​oondream Cloud will shut down, with unspent purchased credits refunded.

That notice went up alongside the reason for it. M​oondream 3 Preview and its successor M​oondream 3.1 are now part of Prime Intellect, which announced on 13 October 2026 that it had acquired M​oondream, with founders Vik Korrapati and Jay Allen joining to lead and scale inference. M​oondream's own post is dated 14 October. This is not a launch and it should not be written about as one. It is an ownership change with a hard migration deadline attached, and if you have a production call to api.moondream.ai anywhere in your stack, you have about four weeks from the date of that post to decide what replaces it.

What Prime Intellect actually bought

The announcement is unusually explicit about the two things that changed hands, and neither of them is a model checkpoint. The first is the team: Korrapati and Allen are joining to run inference and to accelerate multimodal work including vision and computer use. The second is Photon, M​oondream's inference engine, which compiles each model into a single GPU program using custom kernels and a megakernel compiler behind them.

Read the framing and you can see why. Prime's post leads with serving, not with architecture — its case for the deal is that "serving is where a model's usefulness gets decided." The numbers it cites are throughput numbers: M​oondream 3.1 at 34 requests per second, described as roughly 24x faster than the next most accurate open model; more than 60 M​oondream inferences per second on a single H100; answers in tens of milliseconds. Those are Prime's figures and Prime is now the party selling the inference, so treat them as vendor-reported. What is not a vendor claim is the direction of the deal: a company whose product was a vision model bought an inference engine and put its authors in charge of inference.

For anyone holding a M​oondream integration, the practical consequence is that the open weights are staying open and, in M​oondream's words, "will keep improving, now at a much faster rate." What is going away is the convenience layer — the hosted API and Lens, the fine-tuning product — which is exactly the layer a lot of small teams built on because they did not want to run GPUs.

What M​oondream 3 Preview actually is

Worth restating, because a model released in September 2025 that is now being superseded inside an acquisition is a model many teams are only meeting through the shutdown notice. It is a fine-grained sparse mixture-of-experts, 9B total parameters with 2B active per token, initialised from the older 2B dense M​oondream 2 by drop upcycling.

• Architecture — 24 layers, the first four dense and the remainder MoE feed-forward blocks with 64 experts and 8 activated per token, GeGLU, hidden dimension 2048.

• Vision encoder — SigLIP-based, with multi-crop channel concatenation for token-efficient high-resolution input. That tiling detail matters later.

• Context — 32K usable tokens, up from roughly 2K in the previous generation, using a custom SuperBPE tokenizer. The preview post is candid that post-training does not fully exploit the long window yet and that the full 32K is mainly there for people fine-tuning.

• Skills — query, caption, point and detect, all native and all returning structured output. Detection and pointing are the two the launch leaned on hardest.

• Reasoning — hybrid. query runs in reasoning mode by default and can be called with reasoning disabled for cheap, simple questions. M​oondream's framing is that this is visual reasoning with grounding: the reasoning trace references specific regions of the image.

• Licence — the Business Source License 1.1 with an Additional Use Grant, which permits commercial and production use, fine-tuning, quantisation and redistribution, and forbids selling a hosted service that exposes general-purpose M​oondream capabilities.

The last line is the one that quietly shaped the ecosystem, and it is worth being precise about it because M​oondream 3.1 moved to a different instrument entirely — the M​oondream Model License 1.0, effective 7 July 2026. The rule is the same in substance and stated more cleanly: if M​oondream is a component of your product you are allowed, and if access to M​oondream itself is the product you need a separate licence. Self-hosting, edge deployment, customer-controlled VPCs, running it on leased cloud GPUs, fine-tuning it as part of professional services and domain-specific APIs that do not expose general-purpose functionality are all explicitly permitted. Provider-operated general-purpose inference, hosted general-purpose fine-tuning and shared APIs exposing a fine-tuned derivative are the three that require a deal.

That is why the shutdown notice exists in the form it does. M​oondream could not simply hand its hosted endpoint to a third party — the licence's carve-out is precisely about that — so the endpoint closes and the migration goes to Prime, which owns it now.

The two numbers worth distrusting

M​oondream's own benchmark reporting deserves more scrutiny than it usually gets, because the company is more transparent about its method than most and the method is doing a lot of work.

For the preview, the launch chart put M​oondream 3 Preview against GPT-5, Gemini 2.5-Flash and C​laude 4 S​onnet. On the referring-expression detection columns the gap is wide and not plausibly an artefact: RefCOCO at 91.1 for the preview against 75.8 for Gemini 2.5-Flash, 57.2 for GPT-5 and 30.1 for C​laude 4 S​onnet. M​oondream's own caveat is the important part — the frontier models do not support grounding natively, so they were queried through a template, and some preview scores were computed on a 100-question random sample rather than the full benchmark. A templated grounding query against a model with no grounding skill measures prompt engineering as much as it measures the model.

For M​oondream 3.1 the method is more consequential still, and it deserves to be stated plainly: the published benchmark scores were produced by fine-tuning the model on each benchmark's own training split using Lens, M​oondream's fine-tuning product — per-task LoRAs, SFT and RL — and then distilling what those LoRAs learned back into the base weights so the downloadable checkpoint scores well with no tuning at all. The 3.1 post says this outright and even notes that the distilled base lands just behind the individual LoRAs. That is a legitimate and interesting workflow, and it is also not the same thing as a cold out-of-the-box number, because the tasks in question had their train splits fed to the model. If you are comparing 3.1 against a model that never saw those splits, you are comparing a tuned artefact to an untuned one.

None of this makes the numbers wrong. It makes the labelling load-bearing, which is the same standard we apply to any vendor chart.

Your three migration paths

If the hosted endpoint is in your stack, the decision is narrower than it looks.

• Self-host with Photon — Photon remains free and runs locally on N​VIDIA GPUs from Ampere onward and on Apple Silicon Macs, with the same API surface as M​oondream Cloud. pip install moondream, load with local=True, and no API key is needed for the base model. The weights download from Hugging Face on first use. This is the path that preserves your code almost unchanged and moves the cost from per-token to per-GPU-hour.

• Migrate to Prime — the notice says M​oondream and Prime will reach out with next steps and points at Prime for inference, fine-tuning or similar needs, with contact@moondream.ai for urgent problems. Prime is now the only place the hosted path continues, and it will be running on Prime's stack rather than M​oondream's.

• Replace the model — if M​oondream 3 Preview was in your stack because it was a cheap, fast, tiled-image detector rather than because anything about it was specifically M​oondream, the shutdown is a reasonable moment to re-run your own evaluation set rather than defaulting to a like-for-like swap.

The thing to avoid is the fourth path, which is doing nothing until 12 November. Credits are being refunded, so there is no financial reason to keep calling an endpoint that is about to stop answering, and a migration that starts the week before a hard cutoff is a migration where you find out what your grounding prompts actually depended on.

If the model you standardise on ends up being one of the 200-plus in the OrcaRouter catalogue, one part of this gets easier: a single key and a single request shape across them, provider list price passed through with no markup, and automatic failover when an upstream degrades. M​oondream 3 Preview is not in that catalogue and neither is M​oondream 3.1 — both are weights you host yourself or reach through Prime — so for this particular migration the routing layer helps with whatever you route next, not with M​oondream itself. It is worth saying that plainly rather than implying otherwise.

What to watch over the next four weeks

Three things will tell you more than any press release. First, whether Prime publishes an inference replacement for M​oondream Cloud before 13 November or after it; the notice gives no date, and a gap between the shutdown and a replacement service is the outcome that costs users most. Second, whether M​oondream 3.1's download licence stays as written under new ownership — the Model License 1.0's hosted-service restriction is the entire reason the old endpoint has to close rather than transfer, so any revision to it changes what is possible. Third, whether the "much faster rate" of improvement M​oondream promised produces a checkpoint rather than an announcement; the last model release under the old company was 3.1 on 7 July 2026, and the releases since have been Photon versions and speech models.

What is already settled is the date. M​oondream 3 Preview has been generally available on its hosted endpoint since October 2025, it has been superseded by M​oondream 3.1 for three months, and on 13 November 2026 the hosted version of both stops. The weights are not going anywhere. The convenience is.

None of that is hypothetical: 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.