Reka Edge targets physical AI at 7B

This is the kind of model release AI developers should watch closely: not another giant benchmark flex, but a serious attempt to make multimodal agents practical on real devices.

• –Reka says the model uses a ConvNeXt V2 vision encoder plus a 6.4B transformer backbone, and compresses images to just 64 tokens per tile to keep multimodal context cheap
• –The headline comparison is about efficiency, not just quality: roughly 3x fewer visual tokens, 5.46 images/sec throughput, and 0.522s time to first token in its internal tests
• –Reka benchmarks it against Qwen3.5 9B, Cosmos Reason2 8B, and Gemini 3 Pro, positioning Edge as a smaller model that still stays competitive on video understanding, grounding, hallucination resistance, and tool use
• –The deployment story matters as much as the evals: local Hugging Face access, vLLM support, and 4-bit quantization that cuts memory from 13GB to 5GB make this a plausible fit for robotics, XR, and on-device automation
• –The open question is ecosystem traction: if developers trust the model card and the performance claims hold up in the wild, Reka Edge could become a sleeper favorite for multimodal agent builders who cannot afford cloud-heavy vision stacks