Reka Edgeq

Reka EdgeQ, On-Device

← Back to Blog

Sep 16, 2026

Reka EdgeQ, On-Device

Reka EdgeQ, On-Device

Reka Edge is a specialized model for visual perception and physical grounding. It is remarkably efficient on image understanding, video analysis, and object detection while remaining par or better in accuracy against other frontier models. This makes it uniquely suited to real-world deployment on highly constrained devices. With EdgeQ, we compound these advantages with an optimized implementation of Reka Edge for Qualcomm’s Snapdragon 8 Elite, the leading processor in Android phones. This makes it uniquely suited to real-world deployment on highly constrained devices, from phones today to smart glasses, XR headsets, and AI wearables next. It is challenging to retain visual understanding quality and real-time responsiveness within a practical thermal budget, yet EdgeQ delivers state of the art performance in accuracy, speed, and efficiency. We use a Samsung S25 to showcase Qualcomm’s Hexagon NPU below. Accuracy Benchmark Reka Edge EdgeQ Gemma 4 E4B

RefCOCO-A 93.01 88.65 73.93

RefCOCO-B 85.72 77.19 63.10

MLVU 72.31 66.93 32.90

EdgeQ (middle column) is the more capable out of box model implementation for your vision tasks. EdgeQ on-device beats even the full-precision server run of Gemma. On RefCOCO and MLVU, the gold-standard benchmarks for object grounding and video understanding, EdgeQ executed on a Samsung S25 phone outperforms Google’s flagship phone model, Gemma 4 E4B (rightmost column), by up to 34 points, on video understanding. EdgeQ uses quantization and processor-specific optimizations to run on-device with small accuracy loss, shown above. In the table above, Gemma E4B and Reka Edge (leftmost) were run on a server at full precision (FP). This yields a generous best-case accuracy estimate for comparison purposes, even though neither model can run on the S25 with full precision. The speed and efficiency sections use Gemma's only S25-feasible build.

REFCOCO-A / GROUNDING

+14.7

88.65 against 73.93

REFCOCO-B / GROUNDING

+14.1

77.19 against 63.10

MLVU / VIDEO

+34.0

66.93 against 32.90

REFCOCO-A / GROUNDING

+14.7

88.65 against 73.93

REFCOCO-B / GROUNDING

+14.1

77.19 against 63.10

MLVU / VIDEO

+34.0

66.93 against 32.90

Speed Time to first token (s) EdgeQ Gemma 4 E4B

Single image .73 1.6

Video (6 frames) 2.4 6.2

EdgeQ demonstrates real-time responsiveness on vision understanding, beating Gemma’s latency on both single-image and 6-frame video tasks. We use 6-frame clips since that matches MLVU, the gold standard benchmark, and because it is a feasible temporal resolution to capture real-world behaviors. We were able to achieve more than a 2x speedup by customizing the encoder and decoder separately for the S25’s Qualcomm Hexagon NPU. The times shown above reflect latency from the arrival of the image or clip to the first response of the AI model, in other words, the time to first token (TTFT). This time reflects the model-side delay to first detect presence of an action once a clip is captured. Although the E4B is Gemma’s flagship Android phone model, it has no on-device NPU path, its only working route on the S25 is the Adreno GPU (native LiteRT-LM). In order to fit on-device, Gemma needs the quantized .litertlm build. The time to first token indicates the time for vision encoding and the prefill to first token, measured on-device, single-tile frames, warm.

REKA EDGEQ / NPU PATH

Hexagon NPU

AI accelerator

vision + language both on Hexagon NPU

kernels

custom, encoder and decoder tuned separately

GPU during inference

idle, free for graphics

GEMMA 4 E4B / NO NPU PATH

Adreno GPU

AI accelerator

None usable

kernels

LiteRT-LM runtime

GPU during inference

fully occupied

REKA EDGEQ / NPU PATH

Hexagon NPU

AI accelerator

vision + language both on Hexagon NPU

kernels

custom, encoder and decoder tuned separately

GPU during inference

idle, free for graphics

GEMMA 4 E4B / NO NPU PATH

Adreno GPU

AI accelerator

None usable

kernels

LiteRT-LM runtime

GPU during inference

fully occupied

Efficiency Sustained load (240 s) EdgeQ (NPU) Gemma 4 E4B (GPU)

Inferences completed 79 27

Mean inference (s) 3.0 s 8.8 s

Battery power 8.25 W 8.32 W

Skin temp peak 61 °C 71 °C

Latency drift +15% +36%

Energy per inference 25 J 73 J

Both models received identical 6-frame inputs with preprocessing inside the measured time, from a matched cool start, unplugged, over a fixed 240 s window, display at minimum brightness, with all outputs verified correct. The practical advantages of EdgeQ widen during continuous perception, the real physical-AI workload. Running our 6-frame inference test from above, back-to-back for 240 s, Gemma’s model brings the peak phone temperature to 71 °C to the touch (skin-sensor zone). Meanwhile, the GPU junction temperature hits a peak of 105 °C, this thermally throttles the processor, degrading turn latency by 36%. The heat build-up slows sustained inference, and by the final 5-inferences in this sustained load benchmark, the average full inference time was 11.6 s per clip. In practice, applications trigger inference on demand or sample a continuous feed at a fixed rate rather than running back to back. Our test deliberately saturates the processor to rigorously expose each stack's ceiling. What matters for a wearable is heat per inference.

REKA EDGEQ / HEXAGON NPU

6.0 s flat across all 20 turns · peak SoC 39 °C

GEMMA 4 E4B / ADRENO GPU

6.2 s rising to 9.9 s, +58% · peak SoC 103 °C

Twenty back-to-back turns of the 6-frame test, 240 s of continuous vision. Bars are to scale against a 9.9 s maximum; the darker segment is thermal drift.

REKA EDGEQ / HEXAGON NPU

6.0 s flat across all 20 turns · peak SoC 39 °C

GEMMA 4 E4B / ADRENO GPU

6.2 s rising to 9.9 s, +58% · peak SoC 103 °C

Twenty back-to-back turns of the 6-frame test, 240 s of continuous vision. Bars are to scale against a 9.9 s maximum; the darker segment is thermal drift.

On edge devices with limited battery life, EdgeQ can complete about 3x more inferences for the same power consumption over time. A single inference costs about 25 J, a third of the heat of Gemma. For the same test of 240 s of sustained vision, the phone’s case temperature peaked at only 61 °C and the 6-frame latency remained relatively stable with a minimal +15% increase. EdgeQ's implementation keeps the GPU idle, in practice this frees it for what it is specialized for: graphics workloads. This...