Odyssey-3 vs MiniMax H3: The Difference Is a Download

On 3 August 2026 the weights of a 2K video model landed on Hugging Face under a name anyone can pull down and run: MiniMax H3. Six weeks later, on 15 September, Odyssey published a research page for Odyssey-3 — a foundation world model that drives robot arms, a car, a drone and three video games — and put one sentence where the download link would go: it will be released publicly "in the coming weeks." Those two sentences are the entire comparison. One of these models can be in your cluster tonight, on your own silicon, under a licence you can read. The other cannot be bought, rented or downloaded by anyone outside its lab, and the company has not said when that changes. That asymmetry is not a footnote to the matchup. It is the matchup.

What MiniMax H3 actually gives you

H3 was announced on 31 July 2026 and open-sourced on 3 August. It is a general model rather than a single-task video generator: it reads text, images, video and audio as context, and writes video of up to 15 seconds at up to 2K with native stereo audio at 32 kHz and 24 frames per second. Six aspect ratios are supported (21:9 through 9:16), and dialogue is stably supported in eleven languages. There are two published variants of the open base — H3-Base-FL2VA for text-to-video and first-frame, last-frame or first-and-last-frame generation, and H3-Base-Ref2VA, which accepts up to nine images, three video clips and three audio clips, capped at twelve input files.

The part that matters for this comparison is where the open half stops. The full H3 system is a three-stage pipeline: H3-Context-IR preprocesses and orchestrates the context, H3-Base generates at 768p, and H3-Regenerate-2K lifts the result to 2K. Only H3-Base's weights are public, and they ship under the MiniMax H3 Community License rather than Apache or MIT — a licence with conditions attached, which is worth reading before it ends up in a commercial pipeline. The Context-IR stage and the 2K regeneration module run on MiniMax's side. So "open weights" here buys you the middle of the pipeline. A team that downloads H3-Base and self-hosts it gets 768p video generation and no path to 2K without calling the vendor.

What Odyssey-3 gives you instead

What Odyssey-3 gives you instead is a claim and a set of demonstrations. Odyssey describes it as a foundation world model, and the architectural detail it has published is specific enough to be checkable later: an autoregressive diffusion transformer trained on a large collection of visual observations of the world, which the company says yields a learned understanding of physics, dynamics, cause-and-effect and human behaviour. Task adaptation is done by bolting on an action decoder trained on observation-action pairs while the pretrained world model stays frozen — a small policy reads the frozen model's representations rather than fine-tuning the model itself.

The demonstrations span six embodiments. Robot arms, after tens of hours of demonstrations, including recovery behaviours that were not in the training data — reorienting a gripper after a missed grasp. Humanoid policies built with Flexion from tens of hours of teleoperation data, which Odyssey says generalise better than the VLA baselines it tested against, without publishing a number. Closed-loop driving learned from twenty hours of simulated data, driven in India. Indoor obstacle-avoiding drone flight from simulated data. Game environments, including policies trained in Grand Theft Auto V that transferred to Red Dead Redemption 2 and Sleeping Dogs without extra training in those titles; roughly two hours of GTA footage was enough to produce horseback movement in RDR2. And environment generation for training other AIs, tied to the company's PROWL reinforcement-learning work.

The one hard number on the page is a comparison against real footage: Odyssey reports that simulation-trained driving policies travelled about 77% as far between safety-driver interventions as policies trained on real data. That is a vendor figure, unreproduced, with no technical report behind it and no third party has published a result on Odyssey-3. It is the kind of number that decides whether a world model is useful, and it is currently a single company's word.

Two lock-ins that look nothing alike

It is tempting to file MiniMax H3 as "the open one" and Odyssey-3 as "the closed one," but the licensing reality is more interesting than that, and it runs in both directions.

MiniMax H3 is open in the middle and closed at the edges. You can hold the generator, inspect it, fine-tune it, and run it without a network call. You cannot hold the context orchestration that makes it a general model, and you cannot hold the 2K step. If your product's quality claim depends on output resolution, your open-weights deployment has a ceiling that only the vendor's hosted tier lifts.

Odyssey-3 is closed at every layer, and the company's own framing is why that is strange. A world model pitched at robots, drones and vehicles is pitched at buyers who run inference on hardware they control — factory floors, test vehicles, airframes. Those buyers cannot use a hosted endpoint for the control loop; latency and connectivity rule it out. So the business model for a model like this has to involve weights, or it has no customers in the segments Odyssey keeps naming. The absence of a licence today is less a statement of intent than a statement that the commercial terms are still being written. Odyssey's $310 million Series B, announced in June 2026 at a $1.45 billion valuation, with AWS as preferred cloud and Trainium as the compute substrate, is the context in which those terms will be decided.

Dimension by dimension

• What it produces — MiniMax H3: a rendered clip, 4–15s, up to 2K with stereo audio. Odyssey-3: a simulated world state plus motor actions for attached hardware.

• What you can download — MiniMax H3: H3-Base weights, community licence, 768p. Odyssey-3: nothing; no weights, no licence, no repository.

• What stays with the vendor — MiniMax H3: H3-Context-IR and H3-Regenerate-2K, the two modules that make it general and make it 2K. Odyssey-3: everything.

• Price — MiniMax H3: published per-second rates, roughly $0.08 at 768p and $0.13 at 2K, plus $0.05 per second of regenerated output (vendor figures). Odyssey-3: none published anywhere.

• Availability — MiniMax H3: open since 3 August 2026, callable now. Odyssey-3: announced 15 September 2026, "in the coming weeks," no date.

• Third-party evidence — MiniMax H3: ranked first in video editing and second or third in text-to-video and image-to-video on third-party evaluations. Odyssey-3: one vendor figure and no external runs at all.

If you need to build something this month

The practical consequence is that only one of these two is a decision you can act on. MiniMax H3 is routed on OrcaRouter as minimax/minimax-h3 at the provider's list price, with 0% markup — so the vendor's rate is the rate you pay, and when MiniMax moves a number your bill moves with it the same day rather than at renewal. That matters on a per-second video model more than on most things, because the 768p-to-2K spread is nearly two-to-one and the difference between a prototype and a launch can be the difference between those two tiers.

There is a second reason to reach for a router here rather than a direct vendor integration. H3's open half is genuinely useful on its own, and the teams most likely to care about that are the ones with GPUs and opinions about them. Those teams end up in a mixed setup: self-hosted H3-Base for the volume work, a hosted path for the 2K regeneration, and one or two other models for whatever the pipeline needs around the video. One key and automatic failover across providers is worth more in that shape than in a single-model stack, and the routing DSL composes the surrounding models — a prompt model, an image model for the start frame, the video model — into one call instead of three integrations.

Odyssey-3, to be explicit, is not routed by OrcaRouter or by anyone else. There is no endpoint to route to and no price to pass through. This article does not claim otherwise.

The question worth asking both vendors

MiniMax has already answered the interesting question about H3, and the answer is "the middle, at 768p, under our licence." If you are evaluating H3, the thing to test is whether the open base plus your own infrastructure beats the hosted 2K path on cost per usable second — and to test it before you build a product claim on 2K output.

Odyssey has not answered its version yet. When the release window closes, the first question is not a benchmark, it is a licence: which of these six embodiments is a supported deployment, on whose hardware, at what price, and whether the action decoder that makes a frozen world model drive a robot arm is something a customer trains or something Odyssey ships. A world model that reaches robot arms, humanoids, cars, drones and three games with one backbone is a generality claim, and generality is what makes it a foundation model rather than a policy — which is exactly why the commercial terms, not the demonstration reel, will decide who can use it.

Until then the honest scoreboard reads: one model you can pull, one model you can watch. Watch for a repository and a licence first, and a benchmark second — in that order, because without the first two the third is unverifiable.

Put the whole pipeline behind one API key with automatic failover instead of three separate integrations.