AI Potluck
Model components / Base / pretrained models

Reka Core

Reka AI

Reka AI's multimodal foundation model, announced April 2024, natively processing text, images, video and audio with a 128K context and delivered through an API or on-premises and on-device deployment. Reka has since extended the stack with an open-weights Flash release and a multi-agent product, while Core itself remains the closed flagship.

The launch page publishes no usage figure, so adoption is directional. The training-scale details this record once carried are not on that page but in Reka's separate technical report. No 2026 successor to Core itself was found. Verified 2026-08-13 via the Reka Core launch page.

Openness

1 high confidence
1.0
weights
closed
data
closed(mix of public+licensed proprietary)
code
closed
license
Proprietary(API/on-prem, no weights)

Proprietary multimodal model; API/on-prem/on-device deployment but no downloadable weights.

Adoption

2 low confidence
2.0

Distributed through Reka's own API and through Snowflake Cortex and Oracle Cloud Marketplace, but no usage has been disclosed, so level 2 is a directional read rather than a measured one: a 2024 frontier-challenger now well behind. The Snowflake and Oracle partnerships are not enumerated on the launch page cited here.

Capability

2 low confidence
2.0

Frontier-class for early 2024, around the GPT-4V tier, but roughly two years stale against the 2026 anchors, which is why it scores 2. The headline comparisons - matching GPT-4V on MMMU, outperforming Claude 3 Opus on multimodal human evaluation, surpassing Gemini Ultra on video - are Reka's own launch claims and appear on the launch page cited here.

Verified 2026-08-13