Research Engineer - Data Foundations

Deploy Runway's video models on Baseten. You can now deploy Runway's Gen-4.5 and Aleph 2.0 on secure, single-tenant environments on Baseten. Published. October 14, 2026 Baseten has partnered with Runway to bring its video models to enterprises on Baseten. You can deploy Runway's Gen-4.5 and Aleph 2.0, including fine-tuned versions, in single-tenant environments through Baseten Dedicated Inference. Runway builds models that generate and edit cinematic video using text, images, audio, and existing footage. Runway's frontier video generation and editing models, Gen-4.5 and Aleph 2.0, are now available directly through Baseten. Runway's video models: Gen-4.5 and Aleph 2.0. Gen-4.5 turns text and image inputs into video, with support for detailed scene direction, camera movement, and a range of visual styles. You can use it to visualize a storyboard, animate a product image, create cinematic B-roll, or adapt a campaign concept into multiple variations. Aleph 2.0 complements Gen-4.5 by letting you modify generated or real footage you already have. Starting with a prompt or edited keyframe, you can change a specific element or restyle a scene, then carry that edit across relevant shots while preserving the surrounding details. The model supports multi-scene videos up to 30 seconds at 1080p. Runway offers these models through Runway Creative, its platform for generating and editing video, images and audio, and Runway Dev, its API platform for developers. Through this partnership, Runway is also making its proprietary video models available for dedicated deployment on Baseten. Deploy Runway models on Baseten. You can license Runway model weights directly from Runway and deploy them in a secure, single-tenant environment on Baseten. Each deployment runs in an isolated workload plane with dedicated GPU resources, while Baseten manages the control plane and underlying infrastructure. Runway and Baseten support each deployment together, from model customization through production serving. * Security and compliance: Compute and scheduling are isolated to your organization. Baseten's access to Runway's model weights is governed by a distribution agreement, and your deployment runs on infrastructure covered by Baseten's SOC 2 Type II and HIPAA compliance programs. * Fine-tuning: You can fine-tune Runway models on your own data so their outputs reflect your brand, canon, and content guidelines. Your direct license with Runway governs this work, while Baseten provides the infrastructure to run the fine-tuning and deploy the resulting model in a dedicated environment. * Dedicated capacity: You get reserved GPU capacity for video-generation workloads, with autoscaling to handle spikes in demand. Your Runway deployment runs in an isolated workload plane with dedicated GPU nodes, while Baseten manages the control plane. Runway in the enterprise. Amazon, Robinhood, Dolce & Gabbana, and Microsoft are among the enterprise creative teams already using Runway in production. Baseten backs dedicated deployments with a 99.99% uptime SLA. Ready to join them? Learn more and deploy Gen-4.5 and Aleph 2.0 on Baseten dedicated inference.

Runway Project Continuum: inside Runway Labs' early look at an operating system that draws itself. Magnus Corvin Date Published Oct 2, 2026 Benchmarks: Artificial Analysis · updated daily Project Continuum is not something you can download, call, or pay for, and Runway has not said when that changes. What exists on the record as of October 3, 2026 is a demo. Runway Labs published an "early look" at Project Continuum on October 1, and the accompanying walkthrough shows a desktop whose windows, toolbars and side panels are synthesised frame by frame by a video model rather than drawn by a browser or a window manager. It belongs to the same family as Runway Solaris, the Interface World Model Runway introduced on August 31, which generates a software interface as pixels with no HTML, CSS, JavaScript or DOM underneath it. Continuum is that idea scaled from one app to a whole operating-system surface. This is early-signal coverage, and it is worth being blunt about the tier of evidence: every claim here traces to Runway's own posts and its own demo video, and none of it has been reproduced by anyone outside the company. What the demo actually shows. Runway Labs describes four interaction modes, and the demo puts all four on screen in a single session. A Runway designer's walkthrough of the build is the most useful artefact here, because it shows the seams rather than the polish. - Portals - a prompt bar and a strip of generated surfaces. Selecting something already on screen spawns a new pane built around that selection, and the pane is a generated rectangle of video, not a rendered window. The walkthrough keeps a drawer of previously generated portals on the right-hand side, so the session accumulates rather than restarts. - Responsive Video Interfaces - resizing a window is treated as an input, not a redraw. Drag the edge and the seam regenerates with it, which is the clearest single demonstration of the thesis: in a code-rendered UI, a resize is a layout pass; here it is a new generation conditioned on the gesture. - Visual Thinking - an agent mode that works on a visual artefact directly. In the demo it designs a 3D-printable bike mount, iterating on the object on screen rather than describing it in text. - Interactive Worlds - the full-screen case, in the demo a Mars outpost you move through. This is the mode that overlaps most obviously with Runway's general world-model work rather than with interfaces specifically. Runway Labs showed Continuum at the closing of its own Runway AI Summit on September 30 and posted the demo the next day, and the summit's talks page - the surface where the full session is meant to land - is still a notify-me form with no Continuum content on it. The project exists, on Runway's own channels, as a video and a short written description. Two details in the walkthrough are more revealing than the modes themselves. The first is that generated and fixed content are visually distinguishable - the walkthrough points out that the generated surfaces have soft, slightly dreamlike edges where the synthesis boundary falls, which means the system is honest about where the pixels came from. The second is an info affordance that exposes the prompt behind whatever you are looking at. Both are concessions that this is a research application: a shipping product would hide the seams and the plumbing. Where Continuum sits in Runway's research line. Project Continuum is a Runway Labs project, and Runway Labs is the research arm rather than the product line. The vocabulary it is built on was published in August. Runway's framing then was that conventional app development is lossy compression - a designer's intent squeezed into code, then decompressed by a browser, with everything a user can do limited to what a developer pre-programmed. A text-based model that drives a computer still operates inside that constraint: it clicks and types into a UI that someone else fixed in advance. The alternative Runway is arguing for is to generate the interface itself, continuously, as the user acts. Runway calls the family Interface World Models, and Runway Solaris - announced August 31 - is the first member that has been named publicly. Solaris renders interfaces frame by frame, with each interaction conditioning the next frame. There is no code layer to inspect, no DOM to query and no stylesheet to override. Project Continuum is the demonstration of what that feels like across a desktop rather than inside a single app. The third piece of the chain is the September 10 research post, "Towards Instant Video Generation," which is the most technically substantive thing Runway has published in this stretch. It is about making a diffusion video model fast enough to feel live, and it describes the pipeline in three stages: teacher forcing to make the model causal, a student-forcing step to recover the quality that forced causality costs, and then distribution matching distillation run first off-policy and then on-policy, with a curriculum that lengthens the sequences as training proceeds. Runway's stated conclusion is that the inference stack, not the model, was the bottleneck - and that instant generation is the precondition for interactive experiences at all. Read together, the three posts say: here is the thesis, here is what it looks like on a desktop, and here is the machinery that makes it fast. Nothing in the chain is peer-reviewed or benchmarked. The part that is missing. What Continuum is not is the more important half of this piece, because the demo is polished enough to leave an impression the record does not support. - Access - none. There is no sign-up, no waitlist, no request form and no API endpoint. The demo video is the deliverable. - Price - not published. Nothing in Runway's posts discusses cost, seat count, compute requirements or whether a hosted version would be metered per hour or per generation. - Release date - not announced. Runway calls it an early look at a research application and stops there. - Weights - none, and none implied. This is a hosted-research artefact, not an open release. - Vendor page - there isn't one. Runway's own news index carries no Continuum entry; the announcement runs through Runway Labs' social and video channels, and the link from the closing post of the Runway AI Summit resolves to the summit's own talks page, which is still a "get notified when the full talks are available" form. That is a real signal about maturity: a company with a launch to make creates a page first. That list is not a criticism of Runway. It is what an "early look" means. The failure mode to avoid is reading a well-made demo as a product announcement - the demo is a claim about direction, and the direction is the only thing on offer right now. Why this is worth tracking anyway. Two reasons, and neither depends on Continuum shipping in its current form. The first is that the Interface World Model line changes what an application is if it works. If the interface is generated rather than rendered, then the thing you are building against is not an API surface but a conditioning signal - and the tooling, the testing story and the failure modes all move. Runway's own published caveats for this class of system are the interesting part: legible text, factual grounding, consistency across a long session, and accessibility. Those are four hard problems, and none of them is a rendering problem. The second is competitive. A generated interface is a video model doing work that used to belong to a UI framework, and the companies with the strongest video models are the ones best positioned to try it. Continuum is Runway staking a claim on a category before there is a category to compete in. For anyone with a product that would have to call something like this, the practical read is that there is nothing to integrate yet, so the useful work is preparation: know what a real-time interactive endpoint would have to cost, and keep the rest of your stack portable so a swap is a config change rather than a rebuild. That portability argument is the reason OrcaRouter care about it at OrcaRouter - routing through one key across a catalogue of more than 200 models means the model behind a capability is a setting, not an architectural commitment, and a new entrant becomes a route to test rather than a migration. To be exact about its own position: OrcaRouter do not route any Runway model, and Project Continuum is not available through OrcaRouter or anyone else. What OrcaRouter can offer while a category is still forming is a single integration against 0% markup on provider list price, so the price you see for a model is the price its own vendor set rather than a platform's version of it. What to watch. Three concrete tells, in rough order of how much they would settle. - A vendor page. The moment Runway Labs puts a Continuum URL on its own news index instead of on social, the thing has moved from demo to product-track. - A latency number. The whole premise is interactivity. The first published millisecond figure - from Runway or from anyone who gets hands on it - is the claim that matters, and it should be read alongside which hardware it was measured on. - A second interface model. Solaris is one named member of a family Runway said would have several. A successor with published comparisons would make the family a line rather than a single research artefact. Until at least one of those lands, treat Project Continuum as what Runway called it: an early look. The demonstration is genuinely striking, and the record is genuinely thin. Both of those things are true, and the second one is the one to carry into any planning conversation about it.