MiniCPM-V 4.7 35B-A3B vs InternLumina U2: Two Silent Ships, Two Opposite Answers

Neither MiniCPM-V 4.7 35B-A3B nor InternLumina-U2 was announced when its weights appeared. Both are open-weight, both came from Chinese labs, both are the largest multimodal release their team has shipped this year, and both reached the hub with no blog post, no keynote and no launch day. That is where the resemblance stops. MiniCPM-V 4.7 35B-A3B reads images and answers in text; InternLumina-U2 reads images, answers in text, and also draws new ones, edits them, and renders 3D assets. One is built to be small enough to sit on a phone; the other is 16 billion parameters of diffusion research that will not fit anywhere near one. Read side by side, they are not two entries in the same category — they are two different answers to the question of what a vision-language model is for, published within five weeks of each other with nobody watching.

What is knowable here divides cleanly. Everything about the MiniCPM side below comes from the repository, the model card and the config file, all of which anyone can read. Everything on the InternLumina side comes from the lab's own model card and the file tree it has published; the card's own benchmark section is labelled "Preliminary, partial results" and carries no numbers at all. Neither model has been independently evaluated, and nothing in this piece is a launch: OpenBMB did not call MiniCPM-V 4.7 a release until 12 October, and as far as the public record shows Shanghai AI Laboratory has not called InternLumina-U2 one yet.

Two quiet ships, five weeks apart

The MiniCPM-V 4.7 folder first appeared on Hugging Face on 6 October 2026; the hub now reports a creation timestamp of 7 October, and the repository was last touched on 12 October. The InternLumina-U2 repository was created on 1 September. Neither date comes with an announcement attached. But the two have aged differently since, and the difference is the most useful thing in this comparison.

InternLumina-U2 shipped on 1 September as code only. That version was an inference harness, a README and a promise — the weights were "coming soon", the Hugging Face repository held a single .gitattributes file, and the technical report was marked as forthcoming. That is no longer true. The repository now carries two complete hardware-specific checkpoints, one under ascend/ and one under nvidia/, each with seven safetensors shards, its own remote-code module, a parallel plan and a tokenizer. The weights landed quietly and nobody updated the story. As of this writing the repository sits at 16 likes and zero recorded downloads, and the technical report is still not out.

MiniCPM-V 4.7 moved the other way. It arrived in October as weights with no model card — the folder had no README, no licence tag and no benchmarks, which is exactly what our earlier write-up recorded. Ten days later it has all three of the first things and still none of the last: an Apache-2.0 licence, a full card with usage guides, and a claims section with numbers in it, but no evaluation table a reader can quote. So the accurate summary of both is that documentation arrived after the weights did, in both cases, with no announcement in between.

What InternLumina-U2 actually is

The card describes a 16B-parameter mixture-of-experts model with 1B active parameters — written as 16B-A1B — built as a multi-codebook diffusion large language model. The phrase that matters is "fully discrete": InternLumina-U2's central bet is that language and vision should share one token vocabulary rather than living in separate understanding and generation stacks, and that the bottleneck in earlier unified models was the visual vocabulary itself. Its answer is an eight-codebook discrete visual representation built on AToken, which is what lets the same weights both interpret an image and synthesise one.

The scope is broad by design: text QA, text-to-image generation up to 1024×1024, image understanding over natural images and dense charts, instruction-based image editing with an optional dual-CFG mode, video understanding over sampled frames, and 3D understanding over GLB and GLTF assets. The 3D path is the odd one out and the most interesting — assets are rendered into colour-and-depth views through a bundled Blender pipeline before tokenisation, which is a clever way to reuse an image model for geometry without teaching it geometry.

Two engineering details are worth flagging because they are unusual. First, both checkpoints live in one repository, split by accelerator: an Ascend build trained on Huawei NPUs and an NVIDIA build trained on GPUs. That is a deliberate hedge against export and supply constraints, and it means "which checkpoint" is a question InternLumina-U2 users have to answer before anything else. Second, the card states the model is Apache-2.0 with no caveats — genuinely permissive, in contrast to several 2026 multimodal releases that shipped under research-only terms.

What MiniCPM-V 4.7 35B-A3B is now

The card calls it the first MoE in the MiniCPM-V series, and the config file backs that up: 256 experts with 8 active per token, 40 layers, a 262,144-token context window and a 248,144-token vocabulary. Vision runs through a SigLIP2 module with LLaVA-UHD v4 intra-ViT early compression and a mixed 4x/16x token compression mode — the 16x default for throughput, 4x when you need fine detail, selected at call time. It understands single images, multi-image sets and video; it answers in text; it does not generate images.

The claim OpenBMB makes in bold is that this is the first open-source MoE model with lossless, smooth on-device inference, achieved through three named techniques: Lookahead Expert Scheduling, fine-grained layer-wise quantisation, and speculative sampling via the separate DSpark drafts. The number attached to it is 13-plus tokens per second on an iPhone 17 Pro Max. That figure is vendor-reported, unreproduced, and unqualified — no quantisation format, no build, no prompt, no measurement methodology. The card also claims that in Instruct mode the model earns a higher overall score than Qwen3.6 35B-A3B in less inference time, and that claim lives inside a chart image rather than a table, which means it cannot be quoted as a number and should not be treated as one.

What is not a claim is the distribution work, and it is the practical reason to care. Base, FP8, int4, AWQ, GPTQ and GGUF weights all exist; SGLang, vLLM, llama.cpp, Ollama and OpenVINO are supported with day-zero entries; iOS, Android and HarmonyOS apps exist under the MiniCPM-V-Apps repository; and Rockchip, Houmo, Intel, AMD and Arm are all listed as day-zero silicon. Six weeks ago the criticism of this model was that it was an undocumented checkpoint. That criticism is now out of date — it is a documented checkpoint with no benchmarks, which is a different and smaller problem.

The two, dimension by dimension

• What it does with an image — MiniCPM-V 4.7 35B-A3B reads it and answers in text; InternLumina-U2 reads it, answers in text, generates a new one, edits an existing one, and interprets 3D assets.

• Architecture — a 35.2B-parameter autoregressive sparse MoE with 3B active, against a 16B-A1B multi-codebook discrete diffusion model. Different families, not different sizes of one family.

• Modalities in — text, image and video versus text, image, video and 3D.

• Modalities out — text only versus text plus generated images and edited images.

• Weights size — roughly 70 GB of BF16 across 16 shards versus seven shards per hardware build, published twice over.

• Licence — Apache-2.0 on both. This is the one dimension where the two genuinely agree, and it matters, because a permissive licence on a generation model at this scale is not common.

• Deployment story — quantised builds, mobile apps and five inference runtimes versus a remote-code module and a shell script, split by accelerator.

• Public evidence — a card of vendor claims with no evaluation table versus a card that explicitly labels its bench section preliminary and prints nothing in it.

The gap between them is not quality, it is kind

It would be satisfying to declare a winner, and it would be dishonest. There is no shared benchmark between these two models, no third party has run either, and MiniCPM-V 4.7's own evaluation table does not exist while InternLumina-U2's is explicitly marked partial. Anyone ranking them today is ranking documentation quality, not model quality.

What can be compared honestly is the evidence trail, and there the two fail in opposite directions. InternLumina-U2 tells you a lot about architecture and almost nothing about results — a fascinating design with no demonstrated capability. MiniCPM-V 4.7 tells you a lot about results in prose and refuses to publish them as data, which makes the claims hard to falsify and correspondingly hard to trust. The first is a promising instrument; the second is a promising instrument with a marketing layer on top.

The practical read follows from that. If what you need is a model that ingests screenshots, documents or video and returns text, and you want to run it yourself, MiniCPM-V 4.7 is the one with a serving recipe, six runtimes and quantised weights you can download today. If what you need is one model that understands and generates images in the same vocabulary — for editing pipelines, synthetic data, or research into unified token spaces — InternLumina-U2 is the only one of the two that attempts it, and the second checkpoint folder that appeared after the September code drop is the first sign the research is still moving.

What this means if you are just trying to ship something

Both of these are files you download and serve yourself. Neither MiniCPM-V 4.7 35B-A3B nor InternLumina-U2 appears in OrcaRouter's hosted catalogue, so nothing here is a route to calling them as an endpoint, and no article from us should be read as saying otherwise. What a router is useful for in this situation is the other half of the job: the moment your pipeline needs a model you do not want to host. OrcaRouter puts 200-plus hosted models behind one key, passes each provider's list price straight through with no markup of ours, and fails over automatically when a provider degrades — so a team running MiniCPM-V 4.7 locally for bulk document understanding can escalate the hard pages to a hosted model without a second contract or a second SDK.

The nearest hosted model to this shape is Qwen3.6 35B A3B — also 35B total and 3B active, also 262K context, also text, image and video in with text out — available at $0.25 per million input tokens and $1.48 per million output tokens, at cost. That is a genuinely useful control group: if your decision is between self-hosting a 70 GB sparse multimodal MoE and calling an equivalent one, the second option is measurable this afternoon and the first is not.

What would change this comparison

Two documents would settle most of it. A benchmark table from OpenBMB would turn MiniCPM-V 4.7 from a claim into a data point, and the DSpark draft model — the one part of the family with any published numbers, in the form of accepted-tokens-per-step — is the cheapest outside route to a first independent figure. A technical report from Shanghai AI Laboratory would do the same job for InternLumina-U2, and would be the first time anyone has seen its eight-codebook vocabulary scored against anything.

Until then, the honest position is this: one of these two you can deploy, and the other you can only study. That sentence will probably stop being true within a month, and it is the reason neither model deserves to be filed away as vaporware — they are simply earlier in their public life than the writing about them suggests.