Reka AI analysis — onlylabs

Thesis

Reka's public signal has decisively shifted from a general multimodal-LLM / enterprise-automation lab (Reka Flash/Core/Edge, Nexus) toward physical-world AI: 'omni world models' that unify language, images, video, and audio in a single architecture, and a forthcoming 'World Language Action Model' (WLAM) meant to unify reasoning, visual understanding, generation, and action P3 W4 W2. The lab is assembling the full stack behind that thesis: (1) proprietary physical-world data via the Claru marketplace plus open egocentric datasets (RekaDaily-10k, CS2-10k) P6 P8 P20 W1; (2) models spanning on-device vision (Reka Edge/EdgeQ), real-time steerable video generation with NVIDIA, and an inverse-dynamics model that extracts low-level motor commands from video P4 P5 P1 W2; (3) purpose-built evaluation (WorldModelGym, PhysicalRealismBench-U) P13 P28; and (4) commercialization through Reka Vision, Nexus, and enterprise/defense channels P25 P22 P14. Hiring confirms the pivot, with robotics, data-intelligence, GPU-performance, and Claru supply/program roles dominating recent openings E13 E47 E55 P6 P7.

Signal desks

Hiring: Claru data-engine roles — Supply Lead (US Remote) and Program Lead (San Francisco) — are building a two-sided marketplace that sells physical-world data to frontier labs and robotics companies, with a 'team of five' and 'zero to seven figures of ARR in under six months' P6 P7 E9 E10. Technical hiring is broad and distributed: Member of Technical Staff roles for Robotics (UK Remote), ML Engineer (Remote), Product Engineer (US/UK), Applied AI (US/UK), GPU Performance Engineer (US/UK/Singapore), and Data Intelligence (Research, US/UK/Singapore) E13 E41 E44 E54 E55 E47. A Member of Sales Staff (Business Development, US Remote) job points to early GTM, with a JD that flags compute/inference capacity and model serving P10 E15.
Forks: Reka's forks split into two clusters. Agent/multimodal tooling: llama_index, langchain, Archon (agent framework), and diffusers E31 E36 E35 E59. Video/3D/vision research: insert-anything, TAPIP3D (3D point tracking), and Video-Depth-Anything — dependencies consistent with world-model and robotics research E56 E57 E58. Fork metadata is sparse in this pack (parent repo present, but no language/stars for several), so the signal is directional rather than quantitative E56 E57 E58.
Releases: HF model weights: Reka Flash 3 (21B, Apache-2.0), Flash 3.1, Reka Edge (7.1B), and Reka-Inverse-Dynamics-Model (video-classification, Apache-2.0) E1 E2 E3 E4 P1. Open datasets: RekaDaily-10k (10,312 hours egocentric household footage, Apache-2.0) and CS2-10k (10,000+ hours CS2 gameplay with per-frame annotations) P8 W1 P20. Tooling: n8n community node, vLLM plugin for Reka models, and cs2-dem-renderer release tags P16 E26 E48 E51 E39.
Talking: The dominant narrative is physical AI — omni world models, the WLAM roadmap, inverse dynamics for interactive world models, and real-time video generation P3 W4 P2 P5. Reka also publishes heavily on evaluation (WorldModelGym, PhysicalRealismBench-U, Research-Eval) P13 P28 P15, on its data pipeline and datasets P11 P8 P20, and on product framing for Reka Vision (security, broadcast archives), Nexus (AI workforce), and EdgeQ (on-device) P17 P19 P22 P4.

Shipping

Reka shipped a stream of inspectable artifacts in the window. On Hugging Face: Reka Flash 3 (21B, Apache-2.0, 663 downloads / 389 likes), Reka Flash 3.1, Reka Edge 2603 (7.1B params, image-text-to-text), and the Reka Inverse Dynamics Model (RIDM, Apache-2.0) E1 E2 E3 E4 P1. RIDM ships two variants — a flow model (1,795,337 trained params plus 990,162 frozen RAFT-small) and a pixel model (9,836,063 params) — with real-video turn accuracy of 91.5 (flow) vs 51.1 (pixel) P1. Reka also released two large egocentric datasets: RekaDaily-10k (10,312 hours of household footage, roughly 1,670 hours native 4K, Apache-2.0) and CS2-10k (600,000+ player-round videos, 10,000+ hours, per-frame keyboard/mouse/trajectory annotations), plus the cs2-dem-renderer pipeline W1 P8 P20 E22. Developer tooling shipped via the n8n community node, the vLLM plugin for Reka models, and frequent tagged releases of both P16 E26 E48 E51 E39.

Research themes

1. Omni / world models. Reka frames LLMs as an 'intermediate step' toward a single architecture handling discrete and continuous signals that can 'simulate, and reason about the real world' P3. 2. World Language Action Models. Reka team messaging describes training the first WLAM to unify reasoning, visual understanding, generation, and action for egocentric/physical-world applications W4 W2. 3. Interactive world models & inverse dynamics. RIDM extracts low-level motor commands (W/A/S/D, yaw, pitch) from video as 'captions' of physical change, trained on game renders and transferring to real video P2 P1. 4. Real-time video generation. A 30B-parameter real-time steerable video model (720p @ 24fps, 11.8x speedup on a single H100 via NVIDIA FastGen) is positioned as a step toward simulating possible futures P5 W2 W3. 5. World-model evaluation. WorldModelGym measures 'decision-based fidelity' — whether a world model is a faithful simulator for real-world success — complementing PhysicalRealismBench-U, which tests physics understanding via programmatic violations P13 P28. 6. On-device/edge perception. Reka Edge/EdgeQ target spatial grounding, temporal reasoning, and video understanding under constrained thermal budgets on Snapdragon hardware P4 P9.

Hiring & scaling

Hiring maps directly onto the physical-AI stack. The Claru data engine is scaling supply through operator hires who run a paid-collector network (100,000+ collectors per W1) and sell to 'frontier labs and robotics companies' P6 P7 W1. Research/data roles signal a petabyte-scale video pipeline for omni world model training, with a dedicated data team and a Data Intelligence opening P11 E47. Compute/infra hiring (GPU Performance Engineer) and NVIDIA collaboration (FastGen, NIM) underscore inference optimization as a first-class concern E55 W3 P21. Robotics (UK Remote) and Applied AI/Product Engineer roles indicate a buildout toward embodied/action systems and productization E13 E54 E44. Early GTM is thin but real: one sales hire and a Singapore hackathon sponsorship featuring a 'Vice President of Sales and Partnerships, Asia Pacific' P10 E15 P12.

Category implications

Physical AI / world models are becoming the category frontier. Reka's pivot to omni world models and WLAM implies labs are competing on unified multimodal-plus-action architectures, not just LLM scale P3 W4. Monitor WLAM training progress as the next capability marker.
Data is being productized. Claru sells physical-world egocentric data to other frontier labs and robotics companies, and Reka openly releases RekaDaily-10k and CS2-10k — implying data acquisition/collection is now a distinct moat and revenue line in the category P6 P7 P8 P20 W1.
Evaluation is a wedge. Reka ships purpose-built benchmarks (WorldModelGym, PhysicalRealismBench-U, Research-Eval) and cites benchmark deltas (EdgeQ vs Gemma; Reka Edge vs Gemini 3 Pro) — implying evals are both a research scaffold and a GTM differentiator P13 P28 P15 P4 P9.
Edge/on-device deployment is a strategic axis. EdgeQ's Snapdragon 8 Elite optimization and Reka Flash 3's on-device positioning imply real-world deployment on phones/wearables/XR is deliberate product strategy, not an afterthought P4 P27.
Inference infrastructure is an actively managed bottleneck. The NVIDIA FastGen collaboration (11.8x speedup) and GPU-performance hiring imply latency/throughput engineering is core to shipping real-time physical AI W2 W3 E55.
GTM spans enterprise automation, media, security, and defense. Nexus (AI workers), Reka Vision (broadcast archives, smart-home security), and the Oracle Defense Ecosystem show multimodal vision being sold across multiple verticals P22 P19 P17 P14.

Traction highlights

Claru reports 'zero to seven figures of ARR in under six months, with a team of five' and a 100,000+ collector network P6 P7 W1.
Reka Flash 3: 663 HF downloads / 389 likes E1; Reka Edge 2603: 583 downloads / 133 likes E2.
EdgeQ on Snapdragon/Samsung S25 outperforms Gemma 4 E4B by +14.7 (RefCOCO-A), +14.1 (RefCOCO-B), +34.0 (MLVU) P4.
Reka Vision leads recall on Wyze's SmartHome-Bench P17; Reka Edge scores ~93.01 RefCOCO-A vs Gemini 3 Pro's 81.46 P9.
Real-time video model: 30B params, 720p @ 24fps, 11.8x speedup on one H100 W2 W3.
Caveat: external developer/community pull is limited in this pack — HN traction is thin (CS2-10k post at 2 points; Reka Vision post at 5 points) and most developer repos carry low star counts (1–64) E16 E8 E21.