MiniCPM-V 4.7 35B-A3B vs Qwen2.5-Omni: Nineteen Months Later, One of Them Still Hears You

Put MiniCPM-V 4.7 35B-A3B next to Qwen2.5-Omni on a spec sheet and the newer model wins almost every line. It has eight times the context, a sparse architecture that activates a twelfth of its parameters per token, and a decade of inference tooling that did not exist when the older one shipped. Then you get to the audio row, and MiniCPM-V 4.7 35B-A3B has nothing there at all. Qwen2.5-Omni, released in late March 2025, takes audio in and streams natural speech back out; MiniCPM-V 4.7, whose weights appeared in October 2026, takes text, images and video and returns text. Nineteen months of progress in this particular pair did not produce a better omni model. It produced a better vision model that does not do the thing the older model was built around.

That framing is not a knock on either one. It is the whole shape of the comparison, and it is the reason a reader who needs voice and a reader who needs to read a thousand scanned pages should walk away with different answers. On sourcing: Qwen2.5-Omni's figures are the vendor's from March 2025, now nineteen months old and partly validated by wide deployment. MiniCPM-V 4.7 35B-A3B's are OpenBMB's, published on a model card on 12 October 2026 — its first card, ten days after the weights first appeared — and none of them has been reproduced outside the lab.

What the older model actually was

Qwen2.5-Omni was the first end-to-end omni-modal model in that line, and the vendor's framing at the time was explicit: it was the answer to the "zoo of specialists" problem, where transcription, vision and voice each needed a separate model and a separate glue layer. The 7B model took text, images, audio and video as input and replied in text or in streaming natural speech, with two fixed voices, Chelsie and Ethan. It claimed state-of-the-art results on the OmniBench fusion benchmark ahead of Gemini-1.5-Pro, and a speech-generation quality score of 4.51 that the vendor described as human-level.

The engineering held up because it was designed for the constraint rather than around it. Thinker-Talker split the job in two: a Thinker transformer fused the audio and image encoders and produced semantics and text, while a Talker streamed speech tokens from the Thinker's hidden states and shared the full conversational history. Time-aligned multimodal RoPE interleaved video and audio positions so the two streams stayed in sync, and block-wise streaming let the encoders do perception work while the language model handled long sequences — which is what made low-latency spoken replies possible at all.

The limits were just as real, and they were memory rather than context. the vendor's own tables put BF16 inference with FlashAttention-2 at roughly 31 GB of GPU memory for a 15-second video, 42 GB for 30 seconds, and 60 GB for a full minute, against a 32,768-token context window. A minute of video on a 7B model, on one of the largest single GPUs you could rent, is the number to keep in view — it is the number everything after it was built to destroy.

What the new model does instead

MiniCPM-V 4.7 35B-A3B goes the other way entirely. It is OpenBMB's first MoE in the MiniCPM-V series: 256 experts with 8 active per token across 40 layers, a 262,144-token context window and a 248,144-token vocabulary, which comes to roughly 3B active parameters out of 35.2B total. Vision runs through a SigLIP2 module with LLaVA-UHD v4 intra-ViT early compression and a selectable mixed 4x/16x token compression mode — 16x by default, 4x when fine detail is worth the tokens. It handles single images, multi-image sets and video. It answers in text.

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

Seven dimensions, both sides

• Input modalities — Qwen2.5-Omni takes text, images, audio and video; MiniCPM-V 4.7 35B-A3B takes text, images and video, with no audio path at all.

• Output — text plus streaming natural speech in two fixed voices, against text only.

• Context — 32,768 tokens against 262,144, an eight-fold increase.

• Parameters — the 7B figure the vendor names the model by, against 35.2B total with 3B active on the MiniCPM side. The repository itself lists 11B parameters for the checkpoint, because the 7B counts the language model and not the vision and audio encoders sitting alongside it — a useful reminder that "7B" here is a name as much as a measurement.

• Memory footprint — the vendor's own tables show roughly 31 GB for a 15-second video and 60 GB for a minute; MiniCPM-V 4.7 is about 70 GB of BF16 weights before any KV cache, which is why its quantised builds and on-device story carry so much of the card.

• Licence — Apache-2.0 on both, which is the one line where they agree without qualification.

• Deployment tooling — nineteen months of community quantisations and tooling against SGLang, vLLM, llama.cpp, Ollama and OpenVINO support, GGUF/BNB/AWQ/GPTQ/FP8 builds, and iOS, Android and HarmonyOS apps.

• Public evaluation — a nineteen-month-old OmniBench position of 0.561 that has been widely cited, against a card with no quotable evaluation table at all.

The audio hole is the whole decision

If your pipeline needs a model to hear something, this comparison ends in the first paragraph. Qwen2.5-Omni is the only one of the two that takes audio, the only one that speaks, and the only one with an established speech stack — Thinker-Talker, TMRoPE, block-wise streaming, two fixed voices — behind it. It is also the only one of the two anyone has deployed at scale, and nineteen months of community quantisations and tooling is worth more than a spec sheet when you are trying to get a model into an application this quarter.

What MiniCPM-V 4.7 offers in exchange is a much longer memory and a much cheaper per-token compute profile for long visual documents. A 262K window and 3B active parameters is a genuinely different instrument from a 32K window over a 7B dense model — and for a workload that is thousands of page images, screenshots or long video transcripts, the newer model's shape is the right one. It just cannot be the model that answers the phone.

That is the trade, stated plainly: reach against breadth. The older model covers more senses. The newer one covers more page.

Two very different evidence trails

Here the asymmetry runs the other way, and it is worth being exact about.

Qwen2.5-Omni's numbers are old enough to have been partly validated by use. A nineteen-month-old model with a widely cited benchmark position, a decade of downstream quantisations and a large installed base is not a proven quantity in any formal sense, but its claims have had a long time to be contradicted and have not been. The memory figures in particular were published as engineering constraints, not as marketing, and they behaved like engineering constraints.

MiniCPM-V 4.7's numbers have had two weeks and zero outside verification. The model arrived undocumented on 6 October, and the card, the licence and the claims section landed on 12 October. There is still no evaluation table, no technical report, and no third party has published a result. The one place with real published numbers is the separate DSpark draft checkpoint, whose accepted-tokens-per-step figures are OpenBMB's own and which the lab has not published for the parent model.

So the correct posture on the new model is: treat the architecture as read (it is in the config file), treat the distribution as real (the weights exist in six formats), and treat every performance figure as a claim awaiting a second opinion. That last part is not a permanent condition. It is just where things stand on 13 October.

Running either of them, and what the router is for

Both of these are files you serve yourself. Neither MiniCPM-V 4.7 35B-A3B nor Qwen2.5-Omni is in OrcaRouter's hosted catalogue, and nothing here should be read as a claim that we route them — MiniCPM-V 4.7 is a two-week-old checkpoint you would need two accelerators to run unquantised, and Qwen2.5-Omni is a 2025 open-weight model whose practical home is a vLLM instance you operate.

Where a router earns its place is on the other side of that line. Most teams running a self-hosted vision model still need a hosted model for the cases the local one cannot handle — long-horizon reasoning, tool use, the audio inputs MiniCPM-V 4.7 does not accept. OrcaRouter puts 200-plus hosted models behind one key, passes each provider's list price through with no markup of ours, and fails over automatically when a provider degrades, which means adding that second model is a routing change and not a second contract. The nearest hosted shape to the MiniCPM model is Qwen3.6 35B A3B — 35B total, 3B active, 262K context, text-image-video in — at $0.25 per million input tokens and $1.48 per million output tokens, which is a fair yardstick for what self-hosting actually saves you.

Which one, and when

Take Qwen2.5-Omni if you need audio or speech at all, if you want a model with nineteen months of deployment behind it, or if 32K tokens and a 60 GB ceiling for a minute of video fit inside your budget. Its limitations are known and documented, which is a feature when you are shipping.

Take MiniCPM-V 4.7 35B-A3B if your problem is long visual context and text output, if you want the permissive licence, and if you are willing to be early — because early is what you would be. No third party has run it, the on-device claim is unreproduced, and the benchmarks do not exist. What you get in exchange is a documented, quantisable, multi-runtime model with a 262K window and a plausible efficiency story, and you get to be the one who finds out whether the story holds.

What would change the answer: a benchmark table from OpenBMB, or a reproduced acceptance length from the DSpark draft — still the cheapest route to an outside number on this family. What would not change the answer is more time. Nineteen months did not give MiniCPM-V 4.7 the ability to hear, and nothing on the roadmap suggests it is coming.