Reka
Reka develops multimodal AI models and products that work with video, images, text and audio. Its GitHub organization hosts model benchmarks, SDKs and API examples.
GitHub / reka-aiNotable people
Former people
Writing, research & talks
Reka releases RIDM, a compact inverse dynamics model that infers movement and camera commands from consecutive video frames. It is trained only on video games yet transfers to real footage, and the Apache 2.0 weights are on Hugging Face.
Reka previews Rho-1, a 19B omni-reasoning model trained from scratch. It reasons over and generates text, images and video, and emits actions inside a single context, which Reka argues removes the handoff latency of multi-specialist agent pipelines.
- Reka Edgeqreka.ai
Reka ships EdgeQ, an NPU-optimized build of its Reka Edge vision-language model for Qualcomm Snapdragon 8 Elite. It reports RefCOCO and MLVU gains over Gemma 4 E4B, faster first-token latency and steadier thermals under sustained load. Results are vendor-reported.
- Real Time Video Generationreka.ai
Reka and NVIDIA made a 30B-parameter video generation model run in real time: autoregressive, indefinite 720p at 24 fps, steerable with natural-language instructions mid-stream, with an 11.8x speedup that works on H100s. Available in closed beta.
- Real Time Video Generationreka.ai
Reka and NVIDIA report a real-time version of Reka's video model: a 30B-parameter transformer producing an endless autoregressive 720p, 24 fps stream steerable by text instructions, 11.8x faster, running on H100 GPUs. Closed beta.
RekaDaily-10k releases 10,312 hours of commissioned, unscripted first-person household recordings under Apache 2.0, ungated, in a raw tier of full sessions and a captioned tier of short clips, with about 1,670 hours captured in native 4K.
Four flavours of video reasoning, benchmarked: open-vocabulary bounding boxes and masks, timestamp-anchored temporal queries, and identity held across cuts and occlusions by accumulating appearance embeddings, text reads and roster constraints. Sized to run on the camera device itself.
- World Model Data Pipelinereka.ai
How Reka Labs moves petabytes of video from sourcing to training-ready: a formal ingestion spec every dataset must be converted to, ffmpeg metadata and health checks, and the cost-versus-quality tradeoff that decides how deeply anything gets analysed.
- Worldmodelgymreka.ai
WorldModelGym scores world models on decision-based fidelity, asking whether an agent can select the right action sequence by simulating it. A single interface covers latent, predictive, and pixel-based models, and the benchmark deliberately avoids the rollout costs of search-based evaluations.
Deterministic game replays solve a real data problem: Reka renders 600,000+ first-person Counter-Strike 2 rounds at 48fps with per-frame keyboard, mouse-delta, and position annotations, with HUD and weapon artifacts stripped. The renderer that produced the dataset is open-sourced.
Reka releases a physics-violation benchmark built on synthetic scenes with known ground truth, plus an evaluation pipeline. The design requires identifying the offending object and frame range rather than returning a violated/not-violated verdict, which is what separates perception from guessing. Current VLMs fail on that harder task.
Reka Edge and Reka Flash are now served through OpenRouter's gateway, reachable with the OpenAI SDK at model ID rekaai/reka-edge. The post lists the endpoint, key setup, and data-routing controls for teams that want a managed path to a small multimodal model.
Reka's 7B vision-language model pairs a ConvNeXt V2 encoder with a from-scratch transformer backbone and caps output at 64 tokens per image tile. Reka reports state-of-the-art video understanding for its size, lower hallucination rates, and a documented Detect: prompt format returning bounding boxes
Reka's vision API is now an installable n8n community node: natural-language video clipping with burned-in captions, plus image and video Q&A, wired as native workflow actions. Free API key, works on cloud or self-hosted; speech and text capabilities are flagged as next.
Reka Research now samples several reasoning trajectories concurrently and hands them to a separate resolver model trained on verifiable problems. Latency barely moves with N; accuracy climbs on SimpleQA and Research-Eval, exposed as low and high thinking modes with published per-request pricing.
Reka Speech is an 850M transcription and translation model across ten languages, built for bulk offline jobs: a modified vLLM offloads QK embeddings to CPU and recovers word timestamps by dynamic-programming alignment after decoding.
A buildable walkthrough of a search-and-synthesize agent: swap the baseURL on an OpenAI-compatible client, define the answer shape with a Zod schema so results need no parsing, and keep the model's reasoning trace visible. Repo and free API key included.
- End Of Summer Updatesreka.ai
Reka's September batch turns on automatic monthly free credits, ships the Vision API to general availability with docs and code samples, adds two free hours of video upload with unlimited search and QA, and introduces an MCP server for fact-checking workflows.
- Research At Reka Reasoningreka.ai
Reka pushes back on defining reasoning by visible thinking tokens, framing it instead as goal decomposition with dynamic self-verification, and argues that manually orchestrated agent workflows do not generalise the way natively trained reasoning does.
A benchmark aimed at the gap between saturated one-hop factual tests and puzzle-like deep-research suites. Each of its 374 vetted questions carries a checklist of required and forbidden claims for an LLM judge, spreading frontier systems across a 26.7–59.1% accuracy band.
The RL recipe behind Flash 3.1: a DAPO-flavoured REINFORCE variant with dynamic sampling and token-level loss, updated after every rollout to stay on-policy, trained on filtered Numina math and hard coding problems with distributed execution. Released Llama-compatible on Hugging Face.
- Reka Quantization Technologyreka.ai
Reka decomposes low-bit quantization into the primitive and the error-reduction step, then combines LDLQ activation-statistics calibration with full-network self-distillation. The open-source Reka Quant library exports to llama.cpp's Q3_K_S, with a released 3.5-bit Reka Flash 3.1.
Reka Research is a multi-hop web and document research agent orchestrated by a Reka Flash 3.1 model post-trained with reinforcement learning on verifiable rewards. It exposes citation traces, JSON schema output, domain filters and step caps, billed at a flat rate per query.
Reka open-sources Flash 3.1, a 21B model that gains 10 points on LiveCodeBench v5 from reinforcement-learning stack changes, alongside Reka Quant, which uses self-distillation and distributed proxy Hessian LDLQ to cut 3.5-bit quantization loss from 6.8 to 1.6 on average.