Reka AI

Frontier natively multimodal AI models and agents.

Hand-reviewed by the Toolglade team·Updated Aug 2026

chatbots#multimodal#open-weights#agentic-research#video-understanding

Free plan Claimed API Self-hosted

Toolglade’s take

Reka AI is worth a look if you want genuinely native multimodal models, open-weight options, or agentic research without committing to a hyperscaler. Flash 3 under Apache 2.0 and very low entry pricing are real strengths, backed by Nvidia and Snowflake. Temper expectations: it is a small player with much less ecosystem and mindshare than the top labs, its benchmark claims are largely self-reported, and its 2026 pivot toward physical AI and robotics after the Moonvalley merger may shift focus away from the LLM API product.

About Reka AI

Reka AI is a research lab building natively multimodal foundation models that process text, images, audio, and video in one model, plus agentic research and video-understanding products. Its lineup includes Reka Core, Flash and Flash 3, and the compact Edge, with Flash 3 open-sourced under Apache 2.0. Backed by Nvidia and Snowflake, it merged with Moonvalley in 2026 to pivot toward physical AI and world models.

Reka AI builds models that are multimodal by design, handling text, images, audio, and video together rather than bolting modalities on. Its lineup includes the frontier-class Reka Core, the turbo-class Reka Flash and its reasoning-focused Flash 3 iterations, the compact on-device Reka Edge, plus products like Reka Research for agentic deep research and Reka Vision for video understanding. The Nexus platform lets organizations build and manage AI workers. The company was founded in 2022 by researchers from DeepMind, Google, and Meta, and is backed by Nvidia and Snowflake. It open-sourced Reka Flash 3, a roughly 21B reasoning model, under the permissive Apache 2.0 license, and its Reka Quant tooling supports efficient deployment. Pricing is low, with a free hosted option for Flash 3 via third-party hosts. In 2026, Reka merged with video-model startup Moonvalley to advance world models and robotics, signaling a strategic pivot toward physical AI. Earlier acquisition talks with Snowflake, reportedly valuing Reka above $1B, fell through, and the two companies continue to collaborate independently. Reka suits teams wanting cheap or open-weight multimodal models and agentic research without hyperscaler lock-in.

TL;DR

Reka AI is a research lab building natively multimodal models that process text, images, audio, and video in one model, plus agentic research and video-understanding products. It open-sourced Flash 3 under Apache 2.0, is backed by Nvidia and Snowflake, and reached unicorn valuation. In 2026 it merged with Moonvalley to pivot toward physical AI and world models. It suits teams wanting cheap or open-weight multimodal models without hyperscaler lock-in.

Company overview

Reka AI was founded in 2022 by researchers from DeepMind, Google, and Meta, led by CEO Dani Yogatama. It focuses on frontier multimodal AI and, as of 2026, is expanding into physical AI following a merger with Moonvalley.

The company is backed by Nvidia and Snowflake. Earlier talks for Snowflake to acquire Reka, reportedly valuing it above $1B, did not proceed, and the companies continue to collaborate independently.

Product features

Reka's lineup includes the frontier-class Reka Core with 128K context, the turbo-class Reka Flash and reasoning-focused Flash 3, and the compact Reka Edge for on-device use. Reka Research provides agentic deep research with web browsing and code execution, Reka Vision handles video understanding, and Nexus is a platform for building AI workers.

Flash 3 is open-weight under Apache 2.0, and Reka Quant supports efficient deployment. Post-Moonvalley, R&D is shifting toward world models for robotics and physical AI.

Target market

Developers and enterprises needing multimodal understanding such as document, video, and audio analysis and agentic research, cost-sensitive teams wanting open-weight or cheap multimodal inference, and increasingly physical-AI and robotics research.

Buyer personas

End users

Developers and researchers building multimodal understanding and agentic research applications.

Buyers

Engineering and AI leaders seeking cost-effective or open-weight multimodal models outside the hyperscalers.

Key influencers

Open-weight advocates, multimodal and robotics researchers, and cost-focused technical leads.

Ideal customer profile

A team wanting cheap or open-weight native multimodal models for document, video, or audio understanding and agentic research, comfortable with a smaller vendor undergoing a strategic pivot.

Funding & performance

Reka AI has raised roughly $168M cumulatively. Its most recent round was $110M in July 2025, co-led by Nvidia and Snowflake with DST Global participating, pushing its valuation above $1B. A prior Snowflake acquisition attempt did not proceed, and in 2026 Reka merged with Moonvalley to expand into physical AI.

Pros & cons

Pros

Genuinely native multimodal handling of text, image, audio, and video
Open-weight Flash 3 under permissive Apache 2.0
Very low entry pricing plus a free hosted option
Strong research pedigree from ex-DeepMind, Google, and Meta founders
Backed by Nvidia and Snowflake with unicorn-level valuation
Agentic Research and Nexus products for deep-research AI workers
Edge model enables on-device and constrained deployments

Cons

Small player versus the top labs with much less ecosystem
Benchmark claims are largely self-reported
Pivot to physical AI and robotics may divert focus from the LLM API
Model naming is confusing across Core, Flash, Edge, and Flash 3
The failed Snowflake acquisition adds strategic uncertainty
Smaller docs, community, and tooling than hyperscaler platforms
Enterprise pricing is opaque and sales-gated

Pricing plans

Reka Flash 3

~$0.10 input per 1M tokens (free via some hosts)

~21B reasoning model
128K context
Open weights (Apache 2.0)
Free hosted option available

Reka Core

~$2 input / $6 output per 1M tokens

Frontier-class multimodal
128K context
Text, image, audio, video input
Proprietary

Reka Research / Vision

Research ~$25/1K requests; Vision ~$0.05/min

Agentic deep research
Web browsing and code execution
Video understanding
Per-action metering

Enterprise

Custom

Removed rate limits
Monthly invoicing
Volume discounts
Dedicated support

Key features

API

Self-hosted

Multi-language

Integrations

OpenRouter, Snowflake, Nvidia, Hugging Face

Input types

text, image, audio, video

Output types

text

Best For

multimodal understanding, agentic deep research, video and audio analysis, cheap or open-weight inference

Compare key features

View all alternatives →

Feature

Reka AI

MiniMax

Cohere

Pricing

Freemium

Free plan

Yes

Free trial

Yes

API

Yes

Self-hosted

Yes

Team support

Yes

Frequently asked questions

What makes Reka's models multimodal?+

Reka's models are natively multimodal, processing text, images, audio, and video together in a single model rather than combining separate systems, which suits document, video, and audio analysis.

Does Reka offer open-weight models?+

Yes. Reka Flash 3, a roughly 21B reasoning model, was open-sourced under the permissive Apache 2.0 license, free for commercial and research use. Reka Core is not open-weight.

Did Snowflake acquire Reka?+

No. Talks to acquire Reka, reportedly valuing it above $1B, took place but fell through. The companies continue to collaborate independently, and Snowflake remains an investor.

What is the Moonvalley merger?+

In 2026, Reka merged with physical-AI and video startup Moonvalley to advance world models and robotics, signaling a strategic pivot toward physical AI alongside its multimodal LLMs.

Is there a free way to try Reka?+

Yes. Reka Flash 3 is available free through third-party hosts such as OpenRouter, and the open weights can be self-hosted under Apache 2.0.