InternLumina-U2 vs Gemini Omni 1.1 Flash: The Model You Can't Run Yet vs the One You Pay For Per Second

If your deadline is this week, this comparison has a one-sentence answer. Gemini Omni 1.1 Flash is Google's generally-available video generation model — you call it over an API, it returns a clip with synchronized audio, and you pay per second of output. InternLumina-U2 is Shanghai AI Laboratory's quietly published Apache-2.0 research model — you can download its inference code, but not its weights, which the lab still marks "coming soon." One is a product you can buy right now; the other is a paper-shaped bet you cannot run at all. The interesting question is why anyone would put them in the same sentence in the first place, and what each tells you about where open multimodal work is heading in late 2026.

Everything below is labeled by source. InternLumina-U2's details come from the GitHub repository and project page the lab published on 1 September 2026 — the same day the empty Hugging Face stub appeared — and every one of its benchmark figures is vendor-reported, preliminary and unreproduced. Gemini Omni 1.1 Flash's details come from Google's own release materials and pricing page for the model that went generally available on 27 August 2026.

Two answers to the same "multimodal" question

Both models live under the loose banner of "one model, many modalities," and that shared label is the only reason they get compared. Beneath it they have almost nothing in common. Gemini Omni 1.1 Flash is a closed, hosted, video-first generation service: feed it text, an image, a short reference clip or audio, and it produces up to ten seconds of 720p video with speech, music and sound effects baked in, extendable in ten-second increments to a forty-second scene. It reasons about the clip you already have — the extension pass now examines up to ten seconds of prior context instead of the previous one — and it supports first/last-frame control so you can pin the start and end of a shot and let the model fill the middle. It is a tool for making moving pictures, sold by the second.

InternLumina-U2 is a bet in the opposite direction: a single sparse mixture-of-experts model, 16B total parameters with 1B active, that claims to understand images, video and 3D assets and to generate and edit images through one shared discrete-token interface. Its visual side is fully discrete — eight complementary codebooks from a tokenizer the lab calls AToken, so each visual position is described by eight codes rather than one — and its language side is a diffusion backbone the lab extends from LLaDA-2.0. The pitch is that understanding and generation should reinforce each other in a single weight set instead of being stitched together from specialist models. Whether that works is currently unanswerable: the model is not runnable anywhere, because the weights do not exist in public.

The scoreboard, such as it is

The honest scoreboard for this pair has to carry provenance on every row, because one column is a shipping product with published prices and the other is an unreleased research claim with no price at all.

• What it is — Gemini Omni 1.1 Flash: a hosted video generation and editing model (model ID gemini-omni-1.1-flash), GA 27 August 2026. InternLumina-U2: an open-science diffusion LLM for omni-visual understanding, image generation and editing, code released 1 September 2026.

• Weights — Gemini Omni 1.1 Flash: none, closed and hosted-only. InternLumina-U2: none yet either, but Apache-2.0 licensed and explicitly marked "coming soon."

• Output you get — Gemini Omni 1.1 Flash: 10-second 720p clips with audio, extendable to 40 seconds; 1080p and 4K upscales; text alongside. InternLumina-U2: nothing yet — inference code and a roadmap.

• What it takes in — Gemini Omni 1.1 Flash: text, image, video (up to ~131K input tokens; three 10-second clips per prompt), audio. InternLumina-U2: claims text, natural images, dense charts, video over 64 sampled frames, and GLB/GLTF 3D assets rendered to 64 color-and-depth views.

• How you run it — Gemini Omni 1.1 Flash: Google's Gemini API via the stateful Interactions API; consumer access through Google Flow for AI Plus, Pro and Ultra subscribers. InternLumina-U2: self-host only, and only after the weights drop; the code expects a split checkpoint directory, the AToken tokenizer weights, FlashAttention, and Blender 4.5 for the 3D path.

• What it costs — Gemini Omni 1.1 Flash: $0.03/sec at 360p draft, $0.10/sec at 720p, $0.15/sec at 1080p, $0.30/sec at 4K, with a token rate card of $1.50 per million input and $9.00 per million text output. InternLumina-U2: whatever your own GPUs cost, once it exists.

The two columns do not measure the same axis. The Gemini side is priced, served and immediately useful; the InternLumina side is a specification of a model that is not yet a model.

The part that is actually current: Gemini Omni 1.1 Flash's GA

Gemini Omni 1.1 Flash matters on its own terms this week because it is the moment Google's omni video line stopped being a preview. The earlier Gemini Omni Flash preview endpoint is scheduled to be turned off on 30 September 2026, and this GA model is the replacement developers are being moved onto. The upgrade list is concrete: forty-second scene extension built from ten-second increments, keyframe control that lets you specify a first and last frame and have the model generate the connecting footage, reference clips up to three seconds to hold a character or setting consistent, and a 360p draft tier priced at a third of 720p that runs up to 60% faster for iterating on prompts before the expensive final render. If you build video pipelines, the price table — $0.10 for a second of 720p, $1.01 for a ten-second clip, roughly $4 for a forty-second 720p scene built from four extension calls — is the number that changes your cost model.

None of that makes it a competitor to InternLumina-U2 in any operational sense. Gemini Omni 1.1 Flash generates motion; InternLumina-U2, if its claims survive contact with weights, will understand motion and generate stills. A team that needs product video this quarter is not choosing between the two — the Gemini model is the only one of the pair that can be called at all, and it is available through Google's own API and several third-party platforms.

The Gemini API "Models" documentation on Google's AI developer site, captured 2 September 2026 — the page listing the current Gemini family, with the Gemini Omni line in the sidebar navigation.

The part that is not current at all: InternLumina-U2

InternLumina-U2's September 1 release was the quietest kind: three commits to a GitHub repository, a project page, a 28-byte Hugging Face stub whose entire content is an Apache-2.0 license header, and no announcement. What is genuinely public is the inference harness and the architecture, and they are worth a careful read precisely because nobody has been able to test the model itself. The repo shows a driver with one entry point per task — text, text-to-image 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 64 sampled frames, and 3D understanding over Blender-rendered GLB/GLTF views. The design bet is that a multi-codebook discrete visual vocabulary lets one backbone do all of it without inflating sequence length — the same "visual tokens should carry more information" instinct that drives codec-native video models, applied to a unified understand-and-generate interface.

The InternLM/InternLumina-U2 GitHub repository, captured 2 September 2026 — the Apache-2.0 repo landing page with the "Multi-Codebook Diffusion Large Language Model" description, three commits, and the per-task inference scripts that are the entire public release so far.

The benchmark table on the project page is labeled "preliminary, partial results" by the lab itself, and the numbers that are there cut both ways: strong chart and document figures (ChartQA 86.52, CharXiv-DQ 83.65, vendor-reported) sit next to a GenEval of 0.81 that trails the 0.89 of at least one model the lab claims to surpass. There is no independent evaluation, because there is no model to evaluate. Everything about actual quality remains a claim until safetensors files appear in that empty Hugging Face stub.

What a routing layer does when only one side exists

This is one of those comparisons where the routing angle is really about time, not choice. We are not going to pretend OrcaRouter serves InternLumina-U2 — no one can, the weights are not released — and Gemini Omni 1.1 Flash is not in our catalog either; you reach it through Google's own API. What a routing layer is actually for in this gap is keeping your options open without rebuilding your pipeline twice. The pass-through model is the mechanism: when a hosted vendor model does land in a catalog, provider list price is passed through at 0% markup, so the per-second rate the vendor publishes is the per-second rate you pay through one key rather than a per-provider contract. And when an Apache-2.0 weight drop like InternLumina-U2 finally arrives, the rational move is not to rip out your working video stack — it is to stand the new model up on a test path behind automatic failover, so the first time the unproven diffusion model misbehaves, the call falls back to the model you already trust without a code change. Try the new thing where it cannot hurt you; route the thing that pays the bills.

The verdict

If you need video this month, the decision is made for you: Gemini Omni 1.1 Flash is real, priced and generally available, and InternLumina-U2 cannot be called by anyone. If you are watching where open multimodal research goes, InternLumina-U2 is the more interesting artifact — a rare fully-discrete, multi-codebook bet from a major lab, Apache-2.0, with the architecture already public and the weights probably not far behind. Treat the repo as a preview of a research direction, treat the GA video model as a tool you can build on today, and do not let the shared "multimodal" label convince you they are competing for the same job.