Reka Edge 2603

Reka AI's 7B open-weight vision-language model built for on-device agentic use on Jetson, Apple Silicon, and phones.

Reka AI · Open-weight/community models · model

Published in Feed

Overview

Reka AI shipped Reka Edge 2603 on October 9, 2026, a 7B multimodal vision-language model (image/video + text in, text out) optimized for image understanding, video analysis, object detection, and agentic tool use on edge hardware. It encodes a 1024x1024 image in 331 input tokens, versus 1,063 for Cosmos-Reason2 8B and 1,094 for Gemini 3 Pro, and scores 88.40 on VQA-v2, 74.30 on MLVU video understanding, and 93.13 on RefCOCO-A object detection, with 4.69s end-to-end latency versus 16.67s for Gemini 3 Pro. It targets Jetson Thor, Jetson AGX Orin, and 24GB Apple Silicon Macs, with 4-bit quantized builds running on Jetson Orin Nano, Samsung S25, and iPhone at 5GB of memory (down from 13GB) while retaining over 98% of full-precision performance.

This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.

Key features

  • 7B-parameter vision-language model: image/video + text input, text output
  • 331 tokens to encode a 1024x1024 image, vs. 1,063 (Cosmos-Reason2 8B) and 1,094 (Gemini 3 Pro)
  • 88.40 VQA-v2, 74.30 MLVU video understanding, 93.13 RefCOCO-A object detection
  • 4.69s end-to-end latency vs. 16.67s for Gemini 3 Pro
  • 4-bit quantization cuts memory from 13GB to 5GB while keeping 98%+ of original performance
  • Open weights under a custom license: free commercial use only for companies under $1M annual revenue

Use cases

  • On-device robotics and vision perception on Jetson Thor/Orin hardware
  • Local agentic tool-use with image or video understanding on Apple Silicon Macs
  • Mobile vision-language inference on quantized builds (Samsung S25, iPhone)

Pricing

Open weights under a custom license, free for commercial use only below $1M annual company revenue; hosted API priced at $0.10 per million tokens (input and output).

Pricing reflects the source record and may change. Confirm current terms with the provider.

Dates and sources

Released
9 Oct 2026
Record updated
12 Oct 2026
Record last checked
12 Oct 2026

Explore the Feed

Other entries filed under open-weight/community models; inclusion does not imply equivalent features.

  • Muse Spark Open

    Meta's open-weight 816B-parameter MoE model (42B active, 1M context) for coding and long-horizon agent work, released Oct 15, 2026.

  • Nemotron 3.5 Bolt

    NVIDIA's open 74B MoE model (9B active) packed into 24GB for local agents on RTX Spark; released Oct 15, 2026.

  • Jais 3

    Six open Arabic-focused language models from 900M to 375B parameters with 512K context, plus Jais Speech for Arabic recognition.

Browse the Feed