Reka

Reka builds multimodal AI models that process text, images, video, and audio, alongside products for visual search and research. Its releases span the scratch-trained Core, Flash, and Edge family, downloadable reasoning models and quantization tools, and physical AI systems for visual grounding and interactive video generation. The company's early training work provides context for these later products: it built a new GPU training operation with a small team and limited room for experimentation.[5][8][10][12][6][1]

Yi Tay recalled that Reka was underway in late 2022, following collaborations between Google Brain and DeepMind colleagues. In his July 2024 interview, he described cofounder Dani Yogatama initiating the effort and persuading him over several months to join. Tay left Google around the end of March or early April 2023 after wrapping up his PaLM 2 work. He characterized the choice as an opportunity to experience cofounding a company; at Google, he had co-led a PaLM 2 modeling workstream, among multiple co-leads.[1][2]

Reka emerged from stealth on June 27, 2023, announcing $60 million in funding led by DST Global Partners and founding investor Radical Ventures, with participants including Snowflake Ventures. Its research agenda covered general-purpose multimodal and multilingual agents, self-improving AI, and model efficiency. Enterprise AI assistants were then in closed beta. Headquartered in the San Francisco Bay Area, Reka operated as a globally distributed, remote-first team.[2]

According to Tay, much of Reka's compute arrived only in December 2023 after H100 supply delays; for a long period, the company had 500 rented A100s. Acquiring accelerators did not guarantee usable training capacity. His March 2024 infrastructure essay described leased clusters containing hundreds to thousands of chips, with failures ranging from cabling and GPU errors to slow filesystems. He reported that some providers even deleted checkpoints.[1][6]

Reka responded with monitoring, efficient checkpointing, custom scalable storage, and internal workflows for working across clusters. Tay reported that these tools improved utilization and reduced downtime. The team chose PyTorch for accessibility across its members, despite his familiarity with Google's T5X and Mesh TensorFlow. Compute constraints also limited systematic hyperparameter sweeps: Reka used a small number of shorter ablation runs and larger exploratory runs, which Tay called “Yolo runs,” relying on prior training experience to narrow the search.[6]

On February 12, 2024, Reka introduced Flash and Edge in public beta. Flash was a 21-billion-parameter model trained from scratch, positioned for fast workloads; Edge was a compact 7-billion-parameter variant intended for local deployment and latency-sensitive applications. Flash's pretraining covered 32 languages, although deployed multimodal inputs worked best in English at the time. Both base models underwent instruction tuning followed by PPO-based reinforcement learning from human feedback, using Flash as the reward model.[3]

Core launched on April 15, 2024 as Reka's largest model at that time, with a 128K context window and text, image, video, and audio understanding. Reka reported training it from scratch on thousands of GPUs over a few months and offered API, on-premises, and on-device deployment options. The April 18 technical report credited the Reka Team, with 25 named authors. Its reported comparisons were task-specific: Core was competitive with GPT-4V on image question answering and ranked second in a blind third-party multimodal chat evaluation.[4][5]

Latent Space's July 2024 editorial introduction reported 20 Reka employees and fewer than five people working on pretraining. Core's launch named Snowflake, Oracle, and AI Singapore as partners. Snowflake described bringing Core to customers through Cortex; AI Singapore expressed plans to work with Reka's models on the next generation of SEA-LION.[1][4]

On March 10, 2025, Reka released Flash 3 as a 21-billion-parameter research preview with 32K context and downloadable weights under Apache 2.0. Training combined scratch pretraining on publicly accessible and synthetic datasets, instruction tuning, and REINFORCE Leave One-Out (RLOO) reinforcement learning with model-based and rule-based rewards. Its reasoning could be stopped at a chosen budget by forcing the closing </reasoning> tag. The release was primarily English, had not undergone extensive alignment or persona training, and was weaker on knowledge-intensive tasks; Reka recommended coupling it with web search.[7]

The July 10, 2025 Flash 3.1 release added reinforcement learning improvements, a 3.5-bit quantized variant, and Reka Quant. The quantization library supports self-distillation, distributed proxy Hessian computation for LDLQ, and exports to llama.cpp formats including Q3_K and Q4_K. Reka reported a ten-point improvement over Flash 3 on LiveCodeBench v5.[8]

Multimodal Flash 3.1 variants supplied the base models for Reka Research and Reka Vision. Research answers complex questions by browsing the web and private documents; a customized Flash 3.1 orchestrates its work and is optimized end-to-end with Reka's reinforcement learning infrastructure. Vision interprets, searches, and reasons across video and images. By July 22, 2025, both platforms had recently entered general availability, and Reka named Shutterstock and Turing Video as Vision users.[8][9]

On July 22, 2025, Reka announced that it had secured a $110 million investment, backed by new and existing investors including NVIDIA and Snowflake. The company said the investment would accelerate technical development and expand enterprise adoption of its multimodal platforms. It identified its headquarters as Sunnyvale, California.[9]

Reka introduced a new Edge generation on March 11, 2026, a roughly 7-billion-parameter vision-language model for physical grounding, image and video understanding, object detection, and tool use. Its architecture combines a 657-million-parameter ConvNeXt V2 vision encoder with a 6.4-billion-parameter transformer trained from scratch. Encoding each image tile into 64 tokens conserves context and reduces inference work. Detection instructions such as Detect: cups on the table produce object labels and bounding-box coordinates.[10]

RekaDaily-10k, announced August 6, 2026, extended Reka's work into commissioned, unscripted first-person household footage. Reka built Claru as its data engine and reported a paid collection network of more than 100,000 people; a subset contributed household recordings. The September 16 update reported 10,865 raw hours and a processed, captioned tier containing 10,200 hours across 6,373,064 clips, totaling 74.2 TB. Both tiers were fully available under Apache 2.0 and derive from the same recordings.[11]

The recordings preserve the pauses and interruptions of ordinary household activity. Captioning uses the broader session context—for example, connecting loading a washer with returning twenty minutes later to unload it—rather than describing only a sampled frame. The processed tier adds clipping, deduplication, and captions. Reka described collector consent, bystander agreement, and automated privacy screening with human review, while acknowledging imperfect screening and that quality control introduced during collection covered most, but not all, released data.[11]

On August 14, 2026, Reka announced closed-beta real-time video generation, developed in close collaboration with NVIDIA. The system uses a 30-billion-parameter transformer to generate a continuous stream that users can steer with natural-language instructions. Reka reported 720p generation at 24 frames per second on an H100 single-node configuration.[12]

The original diffusion model required more than 100 forward passes per video segment because of iterative denoising and classifier-free guidance. Reka integrated NVIDIA FastGen for timestep distillation, teaching the model to generate in fewer steps and folding guidance into its weights. The company reported an 11.8× end-to-end speedup over its original model without major quality degradation in its blind human evaluation. A proprietary inference engine based on vLLM-Omni coordinates persistent streaming, accepts new instructions, reduces communication and decoding overhead, and applies low-precision quantization. Reka also said it was training a World Language Action Model to unify reasoning, visual understanding, generation, and action.[12]

EdgeQ, announced September 16, 2026, adapted Edge for Qualcomm's Snapdragon 8 Elite and was demonstrated on a Samsung Galaxy S25. Its quantized vision and language kernels target the Hexagon NPU, with the encoder and decoder optimized separately, leaving the GPU available for graphics. Reka's reported warm, single-tile time to first token was 0.73 seconds for an image and 2.4 seconds for a six-frame video clip, including vision encoding and prefill.[13]