Moondream 3 vs MiniCPM-V 4.7: The MoE Storage Bill Nobody Quotes

Both Moondream 3 Preview and MiniCPM-V 4.7 are mixture-of-experts vision-language models, and both are advertised with a small active-parameter number that makes them sound like they fit on modest hardware. Moondream 3 Preview activates 2B of its 9B parameters per token. MiniCPM-V 4.7's headline variant, MiniCPM-V 4.7 35B-A3B, activates about 3B of 35B. Read those two sentences and you would expect the Moondream checkpoint to be the easier one to deploy. Whether it is depends entirely on a question neither launch post leads with: where do the sleeping experts live?

That question is where these two models genuinely diverge, and the answer is MiniCPM-V 4.7's whole reason for existing in the form it does. Its release — MiniCPM-V 4.7 35B-A3B and MiniCPM-V 4.7 1B, from OpenBMB, posted 6 October 2026 — is built around running a 35B MoE smoothly on a phone without degrading it. Moondream 3 Preview, which Moondream released on 18 September 2025 and which is now part of Prime Intellect following the October 2026 acquisition, took the opposite route: keep the MoE small and run it on GPUs with Photon, Moondream's own inference engine. Neither choice is wrong. They are answers to different constraints, and the constraint you are under decides the comparison for you.

Active parameters are a compute number, not a storage number

This is the arithmetic that makes on-device MoE hard, and OpenBMB spells it out in its own model card: an MoE model can have far more total parameters than it activates for a single token, but its on-device storage requirements are not limited to those active parameters. Different tokens select different experts at different layers, and device memory often cannot hold all of them.

• Moondream 3 Preview — 9B total, 2B active. 24 layers, first four dense, the rest MoE feed-forward blocks with 64 experts and 8 activated per token. Hidden dimension 2048, GeGLU inner/gate dimension 1024.

• MiniCPM-V 4.7 35B-A3B — 35B total, roughly 3B active. The first MoE in the MiniCPM-V series. The 1B variant in the same release is dense, not MoE, which matters below.

• The mismatch — divide active by total and Moondream 3 Preview looks like the lighter model by a wide margin. Divide bytes by bytes and MiniCPM-V 4.7 35B-A3B is the one carrying four times the weights.

• Why it does not settle anything — the resident set is what constrains the device, and the resident set is closer to the total than the active count. A 9B MoE does not fit in the memory a 2B dense model fits in, no matter what the router fires.

So "2B active" is a statement about the compute per token. It is not a statement about the memory you need, and treating it as one is the single most common way teams mis-size a MoE deployment.

OpenBMB attacked the storage problem directly

MiniCPM-V 4.7's design is best read as three techniques aimed at exactly that gap. Lookahead Expert Scheduling uses cross-layer routing prediction to preload experts the next layer is likely to want, and prioritises the successfully prefetched ones among those the actual router selected — the prediction guides only prefetching and scheduling, while computation still follows the model's original token-level routing and any required expert the prediction missed is loaded and evaluated as usual. Fine-grained layer-wise quantisation cuts the storage footprint and memory traffic of expert weights. DSpark speculative sampling is integrated into the on-device path to improve decoding efficiency on top of both.

The stated result is 13-plus tokens per second with about 3 GB of memory on an iPhone 17 Pro Max, which OpenBMB describes as the first open-source MoE model running smoothly on-device without compromising performance. Its comparison is explicit and worth noting for what it is: in Instruct mode, MiniCPM-V 4.7 35B-A3B reaches a higher overall score in less inference time than Qwen3.6 35B-A3B. That is a vendor-reported same-size-class claim, and it is the more useful kind because both models are measured on the same harness.

The rest of the deployment story is unusually broad. iOS, Android and HarmonyOS are supported, with day-0 support for chip platforms from Rockchip, Houmo, Intel, AMD and Arm. Inference runs on SGLang, vLLM, llama.cpp and Ollama; fine-tuning on SWIFT and LLaMA-Factory. GGUF, BNB, AWQ and GPTQ quantisations ship, plus FP8. The 1B variant is a separate proposition again — dense, not MoE, and OpenBMB positions it as continuing to improve on MiniCPM-V 4.6 while substantially outperforming Qwen3.5 0.8B on both overall score and inference efficiency, which is the honest comparison for a sub-1B model.

Moondream's answer was to move the hardware, not the weights

Moondream did not try to make a 9B MoE fit on a phone. It kept the architecture and changed where it runs: Photon, its inference engine, compiles each model into a single GPU program with custom kernels, and runs locally on NVIDIA GPUs from Ampere onward and on Apple Silicon Macs, with the same API as Moondream Cloud.

This is a defensible strategy and Prime Intellect's acquisition announcement says why it was attractive: Moondream's team built Photon and the megakernel compiler behind it, and Prime's case for buying the company is that serving, not architecture, is where a model's usefulness is decided. The throughput figures in that post — more than 60 Moondream inferences per second on a single H100, answers in tens of milliseconds, Moondream 3.1 at 34 requests per second — are Prime's own and are now sales material, so label them accordingly. But the direction is unambiguous: Moondream's bet was that you would rather have a fast GPU-backed model than a phone-sized one.

There is a second difference, and it is not in the model card at all. Moondream 3 Preview's hosted endpoint, Moondream Cloud, is being shut down on 13 November 2026 following the acquisition, with unspent purchased credits refunded. MiniCPM-V 4.7 has no such dependency, because it was never primarily a hosted product — it is a set of weights with a broad on-device deployment story attached. If your evaluation included the convenience of calling an endpoint, that convenience is expiring on one side of this comparison.

Task capability: the two cards are honest about different things

Moondream 3 Preview's four native skills are query, caption, point and detect, all returning structured output, and the launch post leads on detection and pointing. The query skill runs in reasoning mode by default, with a documented reasoning=False flag for simple questions, and the reasoning traces are grounded — they reference particular parts of the image, which is why Moondream calls it visual reasoning with grounding rather than reasoning in general. Context is 32K tokens, up from roughly 2K in the previous generation, using a custom SuperBPE tokenizer; the launch post is candid that post-training does not fully exploit the long window and that the full 32K is mostly there for fine-tuning.

MiniCPM-V 4.7's card leads on efficiency and breadth instead: single-image, multi-image and video understanding, with mixed 4x/16x visual token compression that lets you trade visual detail against inference cost per task, and intra-ViT early compression from LLaVA-UHD v4 that reduces computation in the later encoder layers. It also ships a fine-tuning method — Self-Replay Fine-Tuning, aimed at extending a model's capabilities cheaply while reducing forgetting of what it already knew.

• Grounded detection and pointing — Moondream 3 Preview: native skills in the architecture, structured coordinates, grounded reasoning traces. MiniCPM-V 4.7: multi-image and video understanding are the stated strengths; the card does not lead on open-vocabulary detection.

• Video — MiniCPM-V 4.7 supports video input, with a documented example using 128 frames. Moondream 3 Preview's card documents image skills with image-encoding caching for repeated queries on the same image.

• Controllable detail — MiniCPM-V 4.7: a downsample_mode parameter, 16x by default and 4x for finer detail, chosen per call. Moondream 3 Preview: multi-crop channel concatenation in the vision encoder, with the launch post noting the encoder can be tripped up by very small fonts.

• Licence — MiniCPM-V 4.7: Apache 2.0, both variants. Moondream 3 Preview: Business Source License 1.1 with an Additional Use Grant, which permits commercial and production use and forbids offering general-purpose hosted Moondream. Moondream 3.1 moved to the Moondream Model License 1.0 with the same core restriction stated more cleanly.

The licence line is not a footnote for this matchup. If you intend to host the model for other people, Apache 2.0 and Moondream's source-available terms are not remotely the same starting position, and that is true regardless of which one benchmarks better on your task.

Fine-tuning is where both cards point, for opposite reasons

Both vendors tell you the base model is not the product, which is refreshing and also a hint about where the value sits.

Moondream's launch post lists trainability as one of four goals, on the grounds that real vision tasks need specialisation — interpreting an X-ray, spotting struggling people in a crowd — and that even humans need training for those. The company shipped Lens for it, and the Moondream 3.1 launch went further: the published 3.1 benchmark scores were produced by fine-tuning per-task LoRAs on each benchmark's own training split through Lens, then folding what they learned back into the base weights by on-policy distillation so the downloadable checkpoint scores well with no tuning. That is worth reading twice. It is a legitimate workflow and an unusually transparent disclosure, and it also means those numbers are not cold out-of-the-box figures — the tasks in question had their train splits fed to the model.

OpenBMB's equivalent is Self-Replay Fine-Tuning, aimed at a different problem: adding capability without forgetting existing capability, at low cost. Its headline claim is about retention rather than peak score.

If you are going to fine-tune either model, the practical difference is the same one that runs through this whole comparison. Moondream's path assumes a GPU and a hosted service. MiniCPM-V 4.7's assumes you can fine-tune on the frameworks you already run and then deploy to the devices you already target, with a published on-device quantisation story to carry it there.

Where a routing layer earns its place

Neither model here is served through OrcaRouter — both are open-weight checkpoints you host yourself, and one of them is losing its hosted endpoint next month. That makes the interesting question the one both cards dodge: what happens to the requests your chosen model handles badly?

The standard answer is a second model, and the standard cost of a second model is integration: another SDK, another auth scheme, another retry policy. It is the kind of work that quietly consumes a sprint and then becomes permanent maintenance. OrcaRouter's model fusion exists for the narrower version of that problem — a panel of models answering together on the calls where one model's answer is not trustworthy enough alone — and the routing DSL for the broader one, letting you compose several models into a single call so a cheap local model handles the easy majority and something larger takes the residue.

Behind one key there are more than 200 models at provider list price with no markup added, which means a vendor price cut reaches you the same day rather than at the next billing cycle, and failover across providers is automatic when an upstream degrades. For a self-hosted VLM the routing layer sits above the model rather than replacing it. That is the honest description of where it fits in this comparison, and for a system where MiniCPM-V 4.7 35B-A3B is on the device and a hosted model covers the misses, it is the part that keeps the two from becoming two systems.

Which one, decided by the constraint you actually have

If the model has to run on the device, MiniCPM-V 4.7 35B-A3B is the answer and it is not close. OpenBMB built the whole release around the storage problem that stops a 35B MoE from running on a phone — expert scheduling, layer-wise quantisation, speculative decoding — and published the number that follows from it: 13-plus tokens per second in about 3 GB. Nothing in Moondream 3 Preview addresses that problem, because Moondream chose not to have it.

If a GPU is in the picture and grounded detection is the task, Moondream 3 Preview is the more specialised instrument. Native detect and point with structured coordinates and inspectable grounded reasoning traces are a different product from a general multimodal model that also reads text in images, and if a human reviews the output, the trace is the feature.

If you are between the two, do not decide from the benchmark tables. Run your own images through both with your own prompts and score the outputs, because the only figure in this comparison that will survive contact with your workload is the one you produce yourself. And check the licence before the model, because Apache 2.0 and a source-available licence with a hosted-service restriction are a different decision entirely — one of them may end the evaluation before the accuracy numbers matter.

Comparisons like this one are also where the failure modes show up: the documented AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.