MiniCPM-V 4.7 vs Qwen3.8-Omni-Flash: An Hour of Video on Your Phone, or One 1M-Token Call

Two releases, six weeks apart, that both claim to solve video understanding — and they could hardly disagree more about where the work should happen. MiniCPM-V 4.7 35B-A3B from OpenBMB runs on a phone, in about three gigabytes, at a vendor-reported 13-plus tokens per second, and costs nothing per call because the calculator is in your pocket. Qwen3.8-Omni-Flash from A​libaba takes text, image, audio and video into a single one-million-token window, up to a full hour of continuous audio-video per call, and is sold at $0.15 per million input tokens with a headline claim of 98% cheaper audio than its predecessor. One optimises for the device; the other optimises for the context window. For a team with footage to process, that difference decides the architecture before any benchmark does. Qwen3.8-Flash is in the catalogue at $0.15 per million input tokens. Check the routable Qwen models One key, list price passed through at 0% markup, automatic failover across providers. 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.

Two different bets about where the pixels go

MiniCPM-V 4.7 35B-A3B is a sparse mixture-of-experts vision-language model: 35.2B parameters total, about 3B active per token, Apache-2.0, text in and text out, with image, multi-image and video input. It went up on Hugging Face on 6 and 7 October 2026, got its model card on the 12th, and merged into upstream llama.cpp on the 10th. The engineering bet is that most visual work is routine, happens where the camera is, and should never leave the device: mixed 4x/16x visual token compression, expert streaming, and a fork that stacks lookahead expert scheduling, mixed precision and speculative decoding to reach a published 20.51 tokens per second on an iPhone 18 Pro Max.

Qwen3.8-Omni-Flash is A​libaba's native omni-modal model, opened to API customers on 18 September 2026 with the tagline "Omni Senses. Agentic Delivery." One million tokens of context, roughly 991K maximum input, 131K maximum output, a reasoning budget reaching 262K, and up to an hour of continuous audio or audio-video in a single call. Input is text, image, audio and video — including stereo and four-channel spatial audio. Output is text only. It is served on A​libaba's Qianwen platform and A​libaba Cloud Model Studio under the API id qwen3-8-omni-flash; the weights are not being released.

• Where it runs — MiniCPM-V 4.7 35B-A3B: your device, iOS, Android, HarmonyOS, or a desk-side GPU. Qwen3.8-Omni-Flash: A​libaba's API, hosted only.

• Cost shape — MiniCPM-V 4.7 35B-A3B: hardware plus electricity; a 21.2 GB Q4_K_M download. Qwen3.8-Omni-Flash: $0.15 per million input tokens, $0.016 cached, $0.47 output.

• Media per call — MiniCPM-V 4.7 35B-A3B: bounded by device memory and sampling budget, one prompt at a time. Qwen3.8-Omni-Flash: up to one hour of continuous audio-video.

• Output — both text only. Neither model speaks.

• Licence — Apache-2.0 versus a hosted API with no weights.

The 98% is a cost claim; the minutes are the real comparison

A​libaba's headline is that an hour of audio input costs 98% less than on Qwen3.5-Omni-Plus, and an hour of combined audio-video 93% less. That figure is derived by pricing a two-minute sample and multiplying by thirty, with video sampled at 720p and one frame per second — a legitimate way to quote a production workload, and also a description of what the model is being asked to look at. One frame per second is enough to know what was said and roughly what happened. It is nowhere near enough to inspect frame-level operations or catch a single-frame artefact.

Run the same arithmetic against the phone model and the shape of the decision becomes clear. An hour of footage at one frame per second is 3,600 images. On a device decoding at 13 to 20 tokens per second, that is not a one-call job; it is a batched, sampled, summarise-as-you-go job, and the cost is hours of wall-clock time on hardware you already own. Through the A​libaba API, the same hour is one call inside a million-token window, priced in tokens and returned in whatever time the endpoint takes. Neither is free, and they are not the same currency.

The crossover is volume and privacy, not accuracy. A clinic that cannot let frames leave the building has one option. A media-monitoring team with a thousand hours a month has the other, because sampling a thousand hours onto phones is a hardware purchase and a scheduling problem, while a thousand one-hour calls is a line item. Between those poles, the numbers get interesting: the Q​wen price is low enough that a hundred hours of review is a few hundred dollars rather than a few thousand, which is the range where "just send it to the API" stops feeling wasteful.

What each one is actually good at, on the evidence

Qwen3.8-Omni-Flash's launch figures are all against one baseline — Qwen3.5-Omni-Plus — and the average improvement across roughly thirty evaluations is more than 26%. The largest moves are WildClawBench-MM at 71.0, up 36.5 points, and speaker diarisation on AliMeeting, where error rate drops from 88.11 to 3.35 and cpWER from 89.61 to 17.18. A delta that large deserves scrutiny rather than applause; a Chinese technical analysis of the launch attributes much of it to front-end and diarisation engineering wrapped around the model rather than to the model's reasoning, and cautions that the system reads as hearing-specialised with weaker deep video reasoning. That is a critic's reading, not a replication, but it is the right instinct. All of these numbers are A​libaba's own and none has been independently reproduced.

MiniCPM-V 4.7's evidence is thinner in a different way: five evaluation charts, no numeric table, and two directional comparisons — higher overall score in less inference time than Qwen3.6 35B-A3B, and substantial gains over Qwen3.5 0.8B for the 1B sibling. The reproducible facts are the ones in config.json and the file listing: 35.2B total, 256 experts with 8 active, 262,144-token maximum position, a 27-layer vision tower, Apache-2.0, and a 21.2 GB Q4_K_M download. Read the throughput claims as vendor measurements on hardware you have not held, and the accuracy claims as unaudited — which is exactly how OpenBMB presents them.

One thing both share: no independent scoreboard. Neither model appears on a neutral third-party leaderboard with an audited multimodal score at the time of writing. Comparing them on "benchmarks" would therefore mean placing one vendor's chart next to another vendor's chart and calling the result a measurement. Don't.

Agentic sampling versus fixed sampling

The most interesting technical difference is not size or speed — it is who decides what to look at.

Qwen3.8-Omni-Flash ships what A​libaba calls Agentic Understanding mode, in which the model chooses what to inspect rather than processing every frame, gathering evidence in coarse-to-fine rounds. On OmniVideoBench that mode raises accuracy from 63.4 to 67.8 while cutting tokens per query from 145,736 to 79,117 — a 45.7% reduction. Accuracy up and cost down at the same time is the strongest claim in that release, and it is the one that most directly supports the agentic pitch.

MiniCPM-V 4.7's equivalent is a static choice made per image: 16x compression by default, 4x when detail matters. The model does not decide to look closer; the caller does, by passing a different setting. That is the right design for a device with a fixed power budget and the wrong one for a one-hour video where the interesting ninety seconds are unannounced.

The routing question, stated honestly

Neither model is served by OrcaRouter. MiniCPM-V 4.7 35B-A3B is a checkpoint you host; Qwen3.8-Omni-Flash runs on A​libaba's own platforms and is not in our catalogue either. What a routing layer is good for in this architecture is the seam between the two: deciding, per request, whether the clip is routine enough for the local model or long enough to justify a hosted call, and failing over when the hosted side degrades.

OrcaRouter puts 200-plus hosted models behind one key and passes provider list price through at 0% markup, so a vendor price cut — the kind A​libaba just made on audio — reaches your invoice the same day rather than at renewal. It also fails over automatically across providers, which matters more for a hosted omni model than for a phone: the phone does not have an outage, and the API does. One verifiable comparison already sits in our catalogue: Qwen3.8-Flash, the multimodal model this omni variant is built alongside, is routable at $0.15 per million input tokens and $0.47 output — the same list price A​libaba quotes for the new omni model.

Choosing, in one paragraph

Choose MiniCPM-V 4.7 35B-A3B when the frames cannot leave the device, when the workload is many short looks rather than few long ones, when offline operation is a requirement, or when per-call cost is the thing killing your unit economics. Choose Qwen3.8-Omni-Flash when the input is long, the modalities are mixed — audio and video and slides in one call — and you would rather pay per token than build a sampling pipeline. Choose both when you have a phone tier filtering the routine ninety percent and an API tier absorbing the hour-long cases, and accept that you are now maintaining two prompt formats, two sampling policies, and one very clear boundary about which frames are allowed to cross it.