Reka Flash

Σχετικά με Reka AI

Ιδρύθηκε 2022 · San Francisco, California, USA

Reka AI was founded in mid-2022 by Dani Yogatama, Yi Tay, Donovan Ong and Qi Liu, senior researchers from Google DeepMind, Google Brain, Apple and Facebook AI. The company is headquartered in San Francisco with engineering hubs in London and Singapore. Reka raised a $58M Series A in 2023 led by DST Global Partners and a Series B in 2024 at a reported near-unicorn valuation, with backers including Snowflake. Reka's multimodal-by-design family contains three tiers: Edge (~7B), Flash (~21B) and Core (flagship). Reka Flash launched in February 2024 as the cost-performance sweet spot and was the first publicly available Reka model. The Flash technical report described the multimodal training approach later expanded in the April 2024 Reka Core paper. Flash is offered exclusively as a hosted model on the Reka API, Reka Playground and Snowflake Cortex.

Αρχιτεκτονική

Decoder-only Transformer trained multimodally on text, image, video and audio (mid variant)

Reka Flash is a ~21B parameter decoder-only Transformer trained on Reka's multimodal corpus of text, code, images, video frames and audio. As described in the Reka Core technical report, all three Reka sizes share the same architecture (modality encoders projecting into a shared token embedding space) but differ in scale and the proportion of training compute spent on each modality. Flash supports a 128K-token context window, accepts text, image, video (up to several minutes via uniform frame sampling) and audio (via a learned audio encoder), and produces JSON output and tool calls. Public Reka benchmarks place Flash between GPT-3.5 and GPT-4 on MMLU and BIG-Bench Hard, while matching GPT-4V on Perception Test and VideoMME at a fraction of the price. Flash is the recommended default model in Snowflake Cortex's multimodal API and is positioned as the workhorse Reka model for production AI features.

Παράμετροι

21B

Περιεχόμενο

128.000 tokens

Δυνατότητες

• Multimodal input: text, image, video up to several minutes, audio
• 128K-token context window
• JSON output and function calling
• Multilingual coverage across 32+ languages
• Available on Reka API, Reka Playground and Snowflake Cortex
• Strong cost-performance ratio: between GPT-3.5 and GPT-4 quality at much lower price
• 21B parameters give faster serving than Core
• Best for: production multimodal workloads, RAG, audio-grounded chat, video QA

Εκπαίδευση & άδεια

Multimodal pretraining over a curated corpus of text, code, images, video frames and audio with progressive curriculum, same mix as Reka Core at a smaller compute budget than Core but larger than Edge.

Άδεια: Proprietary commercial API. Generated outputs may be used commercially under the Reka terms.

Δοκιμές ασφάλειας: Same safety and bias evaluation framework as the Reka Core technical report; safety filters integrated in the Reka API.