Model

Reka Edge (2026)

OPEN

Reka AI · Reka family · released Mar 11, 2026

New generation of Reka Edge (distinct from the original Feb 2024 7B Reka Edge already in the corpus), positioned for physical AI / robotics / edge cameras; 3x fewer input tokens per image and up to 65% higher throughput than comparable 8B VLMs, with 4-bit quantization cutting memory from 13GB to ~5GB while keeping >98% of full-precision benchmark performance.

ReasoningCodingVisionFunction callingTool useAgentic

1698.3

Elo · unrated

Parameters

Active params

7B (dense)

Context

16K tokens

Architecture

657M ConvNeXt V2 vision encoder + 6.4B dense transformer LM backbone trained from scratch (custom Yasa2ForConditionalGeneration architecture)

License

Reka Edge Community License (free for orgs with <$1M annual revenue; commercial license required above that)

Languages

—

API price (in/out)

$0.1 / $0.1

Modalities

text · vision · video

Benchmark results

Bar shows position within the tracked field; marker = field best

RefCOCOVision93.1%#5

best: Qwen3.5-Omni-Plus · 95.0%

VQAv2Vision88.4%#1

best: this model · 88.4%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Reka family

Elo progression across releases

Reka Edge (7B)Feb 2024378.0 Reka Flash (21B)Feb 2024633.5 Reka CoreApr 2024909.9 Reka Flash 3Mar 20251241.4 Reka Edge (2026)Mar 20261698.3

Resources

HFRekaAI/reka-edge-2603 DOCShttps://reka.ai/news/reka-edge-frontier-level-edge-intelligence-for-physical-ai APIAvailable via Reka AI API

API price $0.1/$0.1 · each benchmark row carries its own source badge (see methodology)