Reka

Almost every well-known AI model — Claude and Gemini included — started as text and had other senses added later. Vision was bolted on, audio was bolted on, and the result works well while quietly betraying its origin: the system reasons in language and translates everything else into it first.

That distinction has consequences for anything involving the physical world. Understanding a video is not reading a description of frames; it involves motion, timing, sound and the relationship between them, and a model that converts all of that into text before thinking about it loses most of what mattered.

Reka's position is that models should be multimodal from the start. Whether that produces measurably better results is the question the company's public materials do not currently answer — but the bet itself is coherent, and the people making it have relevant history.

What Does Reka Build?

• Natively multimodal models — Spark, Edge, Flash and Core — that process video, image, audio and text in one architecture rather than stitching separate models together.
• The stated goal is the physical world, not the chat window: foundational intelligence for things that see and hear rather than assistants that read.
• Video infrastructure is a distinct product: tagging, reasoning over, searching and clipping large volumes of video, available by API, MCP or an app.
• Reka Cloud handles training and deployment of models on its own infrastructure.
• The team is the credential — more than 50 researchers and engineers from Google DeepMind and Meta, including people who worked on Gemini and Veo.

What Does Natively Multimodal Actually Mean?

Reka describes building omni models that process multiple data types within a single architecture for real-time performance, across video, image, audio and text.

The distinction matters and is worth making concrete. A pipeline approach handles video by extracting frames, captioning them with a vision model, transcribing the audio with a speech model, and giving a language model the resulting text. It works, and it throws away almost everything that made it video: the timing between a sound and a movement, the way something accelerates, the fact that two things happened at once.

A natively multimodal model takes the modalities together, so those relationships remain available to it. In principle that is what allows a model to answer questions a pipeline structurally cannot — whether an action looks unsafe, whether a sound came from the thing on screen, whether a sequence happened in the order someone claims.

The real-time framing points at the same target. A pipeline accumulates latency at every stage, which is fatal for anything reacting to the world as it happens.

The honest position for a reader is that this is architecturally sound and, from public materials, unquantified. No benchmarks are published on the site, so how much the approach buys in practice remains an open question you would have to test.

A pipeline that captions frames and transcribes audio has already discarded the thing that made it video: what happened at the same time as what.

Why Four Models Rather Than One?

Reka names Spark, Edge, Flash and Core, and a range like that tells you more about the intended use than any single flagship would.

The naming follows a convention now common across the industry — a small model for constrained environments, mid-sized models trading capability against speed and cost, and a flagship for the hardest work.

Edge is the interesting one given the company's stated focus. A model built to run on a device rather than in a data centre is what robotics, cameras and industrial sensing actually need, because those settings have latency budgets measured in milliseconds and often no reliable network at all.

For a buyer, the practical question is where the boundaries fall: what Edge can do that Spark cannot, what you give up choosing Flash over Core. None of that is published, so it becomes a conversation with the company rather than something you can evaluate in advance.

Video Infrastructure Is The Concrete Product

Alongside the models, Reka offers infrastructure to tag, reason over, search and clip large volumes of video, available via API, MCP or its own app.

This is the most immediately usable thing the company sells, and it addresses a genuinely unsolved problem. Organisations accumulate enormous video archives — security footage, broadcast libraries, recorded operations, inspection footage — and almost none of it is searchable in any meaningful way. Finding the moment something happened means someone watching, which at scale means it never happens.

Search that understands content rather than filenames changes what an archive is, in a way generation tools such as Runway ML do not address at all. Asking for every instance of a particular activity, or every clip where two things co-occur, turns storage into something you can query.

MCP support is a forward-looking inclusion, because it means an AI agent can query the video layer directly rather than through a bespoke integration.

The questions to ask before committing are the ones the site does not answer: how long indexing takes for a given volume, what accuracy looks like on your kind of footage, and — for security and industrial video especially — where the data sits and who can reach it.

• Test on your worst footage, not a demo reel: low light, compression artefacts, poor angles.
• Measure indexing time against a realistic archive size before assuming it scales.
• Establish where video is processed and stored, which for security footage is a governance question.
• Check whether search recall or precision matters more for your use, and test the one you care about.
• Confirm what happens to your index if you stop paying, since re-indexing an archive is not free.

Who Is Actually Behind Reka?

The site states more than 50 researchers and engineers from Google DeepMind and Meta, including pioneers behind Gemini and Veo.

For a research lab this is a more meaningful credential than a customer logo wall would be, and it should be read for what it is. Building frontier multimodal models is not a matter of applying known techniques; it is a small field, and the people who have actually shipped a frontier video model are countable. A team drawn from that group has done the thing before.

It also explains the company's shape. Alongside the models sit Reka Cloud for training and deployment and a data division, CLARU, supplying training data — egocentric video, robotics trajectories, world-model footage. That vertical spread is what a lab builds when it intends to control the whole chain rather than fine-tune someone else's weights.

The data division is a genuine strategic asset. Egocentric video and robotics trajectories are exactly the data that is scarce for physical-world models and cannot simply be scraped, and owning that supply is a durable advantage in a way a model architecture is not.

What the credential does not tell you is what the models can currently do, and no benchmark on the site answers that.

Evaluating Reka Without Published Benchmarks

With no benchmarks, context lengths or pricing published, evaluation has to happen through the company. That is normal for a lab selling to enterprises, and it means going in with your own tests rather than reacting to theirs.

The bar is step three. A natively multimodal model should beat a pipeline on questions about timing and co-occurrence; if it does not, the architectural argument is not paying off for your use.

1. Prepare your own evaluation set before any demo — twenty clips from your real footage with known answers you wrote down first.
2. Include questions a captioning pipeline would fail: what happened at the same time as something else, whether a sound matches the action, what order events occurred in.
3. Run the same set against a conventional pipeline you already use. The gap between them is the entire case for native multimodality.
4. Ask directly for context length, latency figures and the capability boundaries between Spark, Edge, Flash and Core, since none is published.
5. If you need on-device inference, establish exactly what hardware Edge runs on and what it gives up against Flash.
6. Get data residency, retention and training-use terms in writing before any real footage leaves your network.

What Reka Does Not Publish

It is worth being direct about the gaps, because they are the reason this listing scores where it does rather than higher.

No context lengths are stated for any model. No benchmarks or evaluation results appear on the site. There is no statement about whether any weights are open or self-hostable, which for a company targeting edge and robotics is a question buyers will certainly ask. No pricing exists in public. No customer names appear. No funding detail is given.

None of that is unusual for a research lab at this stage, and several of these are ordinarily settled in a sales conversation. But the aggregate effect is that a prospective buyer can read the entire site and still not know what they would be getting.

There is a Trust Center in the footer alongside terms, privacy and responsible AI pages, which is where security and compliance answers would live — worth reading before a demo rather than after.

The practical consequence is that Reka is currently a company you evaluate by talking to, not by reading. For an enterprise procurement that is manageable; for a developer wanting to try something this afternoon, it is a closed door.

Multimodal Model Vocabulary

Four terms that explain what Reka is claiming and why it would matter.

Natively multimodal

A single model architecture trained across modalities together, rather than separate models chained by text. Preserves timing and co-occurrence, which pipelines discard.

Omni model

Reka's term for a model handling video, image, audio and text in one system. In practice a synonym for natively multimodal, used to distinguish from vision-plus-language models.

Edge inference

Running a model on a device rather than in a data centre. Necessary where latency budgets are milliseconds or the network is unreliable — robotics, cameras, vehicles.

Egocentric video

Footage captured from a first-person viewpoint, as a person or robot sees it. Scarce, hard to scrape, and among the most valuable training data for physical-world models.

Reka's Model Line, Listed