If your team calls Moondream's hosted vision API or trained a model in Lens, you have until November 13, 2026 to move it. Prime Intellect has acquired Moondream, and Moondream's announcement says Moondream Cloud, covering both inference and Lens, shuts down that day, 30 days after the news. The base model is the easy part: the weights stay open and run free on your own GPUs. The fine-tunes are the hard part, because Moondream's own documentation says running one locally still needs a Moondream API key, and neither company has said what that key does after the cloud goes dark.
Terms were not disclosed. Prime Intellect's post, dated October 13, says co-founders Vik Korrapati and Jay Allen join to lead inference and work on multimodal models, and that Moondream stays on as the brand for small, efficient multimodal models.
What Prime Intellect Actually Bought
Prime Intellect bought an inference engine and the team that built it, and the hosted product is what gets cut. Photon, Moondream's engine, "compiles each model into a single GPU program," according to Prime's announcement, and Prime is folding Photon and Moondream's megakernel compiler into its own stack so it can serve models other than Moondream's.
Crypto Briefing's write-up frames the deal the same way: Prime is after the inference optimisation software, with the aim of faster inference at lower compute cost. It reports that Moondream had fewer than 10 employees and had raised $4.5 million, and that Prime closed a $130 million Series A at a $1 billion valuation in July 2026 and reported an annualised revenue run rate above $100 million by then. No closing timeline or integration plan has been published.
The model line is not being abandoned. Moondream's post says "Moondream models will remain open, and will keep improving." Prime says Moondream 3.1 is already available through Prime Inference on Photon. Its performance figures are vendor claims: 34 requests per second, about 24x faster than the next most accurate open model, and more than 60 inferences per second on a single H100.
Prime Intellect gets a fast vision model and a compiler it can point at other open models. What you have to work out is which parts of your setup survive November 13.
Which Moondream Workloads Survive November 13?
Base-model inference survives if you self-host it. Hosted inference and anything that depends on Lens has to move. The split looks like this:
| What you use today | After November 13 | What you do |
|---|---|---|
| Moondream Cloud API, base model | Gone | Self-host with Photon, or move to Prime Inference |
| Photon running base Moondream 3.1 locally | Keeps working; no key required since Photon 1.3.0 | Pin the package and weights version |
| Lens fine-tune called through the cloud API | Gone | Migrate to Prime or retrain elsewhere |
| Lens fine-tune run locally through Photon | Unclear: needs MOONDREAM_API_KEY |
Get a written answer before the cutoff |
| Prepaid credits | Unspent purchased credits refunded | Claim the refund and the Prime credits |
The base model is safe because of a change Moondream made in June. The Photon 1.3.0 release note, dated June 8, 2026, says local use is "totally free" with "No API key required." The Moondream 3.1 model card confirms the setup: pip install moondream, a local flag, and weights pulled from Hugging Face on first use, on NVIDIA Ampere or newer GPUs or Apple Silicon. The Photon page lists support from Jetson Orin Nano up to B200.
The same release note carries the catch. You "will still need to pass your MOONDREAM_API_KEY if you want to run a finetuned model." The model card says the same: a key is needed for fine-tunes or for Moondream Cloud.
Why Are Lens Fine-Tunes the Real Risk?
Lens fine-tunes are the real risk because the documented ways to manage them all run through Moondream's servers, and none of them is an export. Lens is Moondream's hosted fine-tuning service, launched April 20, 2026, offering supervised fine-tuning and reinforcement learning through an API with no hardware to manage. The Python SDK documentation describes fine-tunes as LoRA adapters at rank 8, 16, 24 or 32, and checkpoints as server-side objects you save, list and delete by training step. "Only saved checkpoints can be used for inference." There is no download method in that list.
The Lens launch post does say fine-tuned models can run locally with Photon, which suggests the adapter can reach your hardware. Combine that with the key requirement and the likely mechanism is that Photon fetches or authorises your adapter using your account. Moondream does not document that mechanism. Neither the Moondream announcement nor Prime's post says whether those keys keep validating, whether fine-tunes move to Prime's platform intact, or whether customers can download their LoRA weights.
Moondream says it will contact customers directly and lists contact@moondream.ai for urgent issues. If a fine-tune is in production, that email should go out this week, and you want the answer in writing.
The good news is that retraining is cheap. Moondream's own Lens examples put a basketball ball-handler detector at 54 minutes and $16.89 of training, and a glaucoma staging model at 47 minutes and $15.68, per the Lens launch post. The pricing page lists Lens training at $0.60 per million training tokens. The expensive pieces are the labelled images, the reward function and the eval set, so if those live only inside a Moondream workspace, copy them out first.
What Does the Moondream Licence Let You Do Now?
The Moondream Model License 1.0 lets you self-host, fine-tune and ship Moondream inside your own products, including on rented GPUs, and stops you from reselling it as a general-purpose hosted model. The licence text, effective July 7, 2026, grants a perpetual, worldwide, royalty-free right to commercial use, internal use, modification, quantisation and deployment on customer-controlled cloud, VPC, edge or leased neocloud infrastructure. The model card lists Moondream 3.1 under this licence.
The restriction matters if you are a platform team. You may not offer a "General-Purpose Hosted Model Service" without a separate commercial licence. That covers OpenAI-compatible endpoints, per-customer instances where model access is the product, and splitting general access into task-specific services to get around the rule. An internal shared vision endpoint for your own employees is customer-controlled deployment. Exposing that same endpoint to your customers as a generic image API is not. Ask legal to read Section 3 before you build a shared service.
The licence terminates automatically on a violation, with reinstatement if you cure within 30 days of notice. It is a source-available licence with a managed-service carve-out. Check which licence your weights carry, because License 1.0 covers only model files that expressly reference it. If you run an older Moondream release, read the licence file in that repository.
Should You Migrate to Prime or Self-Host?
Self-host if you run base-model inference at steady volume and already operate GPUs. Move to Prime if you depend on hosted fine-tuning and don't want to rebuild that workflow in a month. The licence allows either, and the cost gap is small at low volume: Moondream Cloud lists Moondream 3.1 at $0.30 per million input tokens and $1.00 per million output tokens in real time, half that in batch, per the pricing page.
Prime says every Moondream Cloud customer will receive credits on the Prime platform, usable for Moondream models and its wider training and inference products. Moondream separately says unspent purchased credits will be refunded. Prime has not published Moondream 3.1 inference pricing in the announcement, so you can't yet compare the two on cost.
The steel-man for moving to Prime is real. The team that built Photon now works there, and Prime is a $1 billion company with the capital to keep the model line going, according to Crypto Briefing. The counterpoint: your vision pipeline would then depend on a vendor whose post-merger pricing, data terms and fine-tuning product you have not seen. A hosted vendor with unpublished terms is the situation that produced this 30-day migration.
The pattern is familiar. Cloudflare gave Deno Deploy six months, and Fetcher's customers got nine days. Thirty days sits between those, and it is enough to self-host a base model. It is tight for an untested fine-tune migration.
What to Do Before November 13
This Week:
- Inventory every call to the Moondream Cloud endpoint, every service holding a
MOONDREAM_API_KEY, and every Lens fine-tune and checkpoint ID in use. Include batch jobs and edge devices, which are the ones that fail silently. - Email contact@moondream.ai for each production fine-tune and ask three questions: can you download the LoRA weights, will keys validate for local Photon after November 13, and do fine-tunes transfer to Prime with their checkpoint IDs.
- Copy your training images, labels, reward functions and eval sets out of any Moondream workspace into storage you own.
This Month:
- Stand up Photon with base Moondream 3.1 on one GPU you already have, pin the
moondreampackage and weights version, and confirm it runs with the network cut. Run your eval set against it. - If you can't get fine-tune weights out, retrain the adapter on another stack using your own data, and compare it with the Lens version on the same eval set before the cutoff.
- Have legal read Section 3 of the Moondream Model License 1.0 if any internal vision endpoint is, or might become, available to customers.
Before the Cutoff:
- Claim the refund on unspent purchased credits and the Prime platform credits, in writing, with amounts.
- Cut over, then revoke every Moondream API key so nothing is quietly still calling a dead endpoint on November 14.
The Bottom Line
This is a smaller version of what happens to most hosted AI tooling bought for its engine. The acquirer keeps the technology and the people and closes the storefront. Moondream's open licence keeps the base model safe. Your fine-tunes are covered by nothing in writing, and the inventory you do this week shows how many you have running.
Get the answer about fine-tune weights in writing before November 13.
Continue Reading
- Cloudflare's Deno Deal Gives Deno Deploy Six Months to Live
- Fetcher Shuts Down October 16 After Juicebox Buys Its Assets
- Nvidia Courts Reflection AI Before Beam's Apache Weights Ship
- Microchip's No-EOL Pledge Stops Short of Hailo's Compiler
- Best Air-Gapped LLM Stack: Apache Weights on vLLM Beat NVIDIA's Fee
- FP8 vs NVFP4 vs INT4-AWQ: AWQ Outran NVFP4 on 3 of 4 Models