Deploying physical AI on wearables and phones requires real-time perception inside extreme thermal and battery limits. Reka EdgeQ is an optimized on-device VLM running natively on @Qualcomm #Snapdragon 8 Elite’s Hexagon NPU. ⚡️ 0.73s Time to First Token (Image) 📹 +34 pts over Gemma 4 E4B on MLVU video benchmarks 🔋 6.9 mWh energy per inference (~3x lower heat overhead) By custom-tuning our ConvNeXt V2 vision encoder and decoder directly for the #NPU, EdgeQ keeps the GPU completely idle, minimizing thermal throttling under sustained continuous load. 🔗Our technical breakdown: reka.ai/labs/research/reka-e… #Multimodal #ComputerVision #VLM

VLMs: language, image → language. Omni: language, image, video, actions → language, image, video, actions. 🚪 Same doors in. It is the way out that changes. New #AIresearch from Reka Labs on omni-world models shares why unifying VLA and world-model (WAM) approaches under one backbone matters for physical-world intelligence. reka.ai/labs/research/evolut… #MultimodalAI #WorldModels #PhysicalAI

📊 Update: RekaDaily-10k is now complete. The raw tier finished at 10,865 hours. The processed and captioned tier is live for the first time: 10,200 hours, 6,373,064 clips, 74.2 TB. Both tiers are on Hugging Face now, fully open under Apache 2.0. 🔗 Read the updated post: reka.ai/labs/research/rekada… 🔗 Get the dataset on Hugging Face 🤗: huggingface.co/datasets/Reka… #PhysicalAI #WorldModels #OpenSource #EgocentricData

Together with @nvidia , we've built a real-time version of our video generation model: 30B parameters, steerable mid-stream in natural language, 720p at 24fps. 11.8x faster on a single H100, no major quality drop in our blind eval. We're one step closer to World Language Action Models: AI that doesn't just generate the physical world, it acts within it. Closed beta open now, contact us for access. reka.ai/labs/research/real-t… #RekaLabs #NVIDIA #AIResearch #VideoGeneration #WorldModels

Enterprise data is multimodal. Most AI tools aren't. Reka + @HPE + @nvidia built a validated stack to close that gap, from pilot to production. Here's our solution: hpe.com/psnow/doc/a00161604e… #Reka

we've made freely available a few things over the past few weeks. training datasets - RekaCS2-10k: bit.ly/4zbjpmB - RekaDaily-10k: bit.ly/4gya2F5 eval benchmarks - WorldModelGym: bit.ly/3UvYYR7 - PhysicalRealismBench-U: bit.ly/3Swj512 and share insights into our data processing pipeline: bit.ly/3TZe5Cw we intend to continue sharing our research broadly. the next one is our omni model. still training for now, but looking forward to open weighting it.

To train physical AI, synthetic environments and polished, tripod-mounted footage aren't enough. You need the actual mess of the real world. 🚀Releasing RekaDaily-10k: 10,312 hours of unscripted, first-person household footage as collected, recorded by paid collectors in real homes across the US, LatAm, Asia, and Africa, with roughly 1,670 hours in native 4K. Raw tier available now on @huggingface🤗, full set live by early next week. Apache 2.0. huggingface.co/datasets/Reka… Read more reka.ai/news/rekadaily-10k-e… This is what Claru, our data engine, produces daily from 100,000+ paid collectors. #RekaDaily10k #PhysicalAI

Can AI actually reason about a video, or is it still just listing what's in frame? We gave our model a scenario: twenty competitors, nearly identical vehicles, fast-paced motorsport. It has to figure out who's who from pixels alone. Result: 90.9% accuracy, up from 39.6%. 📄 Read the full research: reka.ai/labs/research/beyond…

A single camera frame might capture a thousand people on a train platform. It's not enough to recognize them. The model has to know which one matters. In our latest post, we share how our models reason about real-world video: spatial grounding, temporal tracking, identity, and judgment. One benchmark jumped from 39.6% to 90.9% top-1 accuracy identifying individual competitors in motorsport broadcasts. The gain came from better evaluation, not a bigger model. 📄 Read more: reka.ai/news/beyond-recognit…

Reka has signed the Open Weights and American AI Leadership letter. We believe an open AI ecosystem is essential to advancing research, expanding access, and building a more innovative and resilient field. That's why we've openly released datasets like CS2-10k and Research-Eval, models like Reka Flash 3 and Reka Edge, and frameworks like Reka Quant. microsoft.com/en-us/corporat…

Today we're opening the WorldModelGym leaderboard 📊 Most world model evals ask one question: does the generated video look real? WMGym asks a different one. Give a model a menu of possible actions, can it correctly predict which one actually leads to the best outcome?🎯 We ran our own Dreamer-v3 across all four benchmark families as the first entry. It currently leads three of them — Meta-World, DeepMind Control, and Classical Control (where it hits 84% decision fidelity), and sits second on Atari. We invite you to also enter your own model: you host a /score endpoint, we call it and compute the score ourselves. Read our full breakdown and methodology: wm-gym.labs.reka.ai/

Today Roberto and Marianna walk through PhysicalRealismBench-U — how the benchmark works, how they built it, and what VLMs still get wrong about basic physics. 🔗 link.reka.ai/labs/physical-r… #PhysicalAI #AI #RekaLabs #ComputerVision #VLM #PhysicalRealismBench

In this video, @MateuszOnAI and @zaheri_h_r sat down to talk through two questions driving recent Reka Labs research. 1️⃣ Can VLMs understand physics the way we do? 2️⃣ And if an agent uses a world model to choose between actions, does it pick the right one? Two benchmarks, two complementary questions about what it actually means for a model to understand the physical world. Read more on the benchmarks here: 📄 PhysicalRealismBench-U → link.reka.ai/labs/physical-r… 📄 WorldModelGym → link.reka.ai/labs/world-mode…