EdgeWise
An edge-cloud selective delegation architecture for agent inference, in which a small edge-side step selector determines which reasoning steps require processing by a large cloud model while the remaining steps are completed at the edge, thereby reducing cost and latency.
Timeline (last 90 days)
EdgeWise is an edge-first long-horizon Agent solution that uses a step selector to determine whether the current reasoning step needs a cloud large model, thereby compressing cloud calls to 10-20% of total steps.
EdgeWise's selector is a 12M-parameter BERT-style binary classification model that runs inference in under 5ms on Jetson Orin and under 3ms on Apple M-series chips.
In a 50-step navigation task, only 12% of decisions required calling a GPT-4-turbo-level large model, and the remaining steps could be completed at low cost on the edge.
Across evaluations in three domains—robotic manipulation, autonomous navigation, and text-to-video generation—EdgeWise delegated 10-20% of steps to the cloud, maintaining over 95% of original performance with roughly an 85% cost reduction.
EdgeWise reported a 95th percentile latency of 190ms for edge-side steps, a cloud call peak of 180ms, and end-to-end latency below 200ms for 94% of steps.
EdgeWise provides an Apache 2.0-licensed 12M-parameter pretrained selector hosted on HuggingFace, as well as reference Docker images for edge and cloud.
Under the traditional all-cloud architecture, each picking action consumes 1200 tokens, costing $0.0096; after EdgeWise selects delegation, each round consumes 300 tokens, costing $0.0012.
The EdgeWise roadmap shows: public Beta in Q4 2026, SDK v1.0 release in Q1 2027, and a broader cloud orchestration service in Q2 2027.
The EdgeWise selector is evaluated using ranking metrics such as MRR (Mean Reciprocal Rank); when MRR approaches 1, it can precisely route key steps to the cloud.
EdgeWise chooses the open-source ONNX Runtime inference engine, letting partners compile the selector themselves on platforms such as Linux, RT-OS, iOS, and custom ASICs.
2 more timeline events
All Facts (12)
A fleet of 100 robots using EdgeWise saved about $45k in cloud spending per month, and the pilot robot completed 10,000 picks within two weeks, with an error rate on par with the all-cloud baseline.
50%UnverifiedEdgeWise's selector is a 12M-parameter BERT-style binary classification model that runs inference in under 5ms on Jetson Orin and under 3ms on Apple M-series chips.
50%UnverifiedIn a 50-step navigation task, only 12% of decisions required calling a GPT-4-turbo-level large model, and the remaining steps could be completed at low cost on the edge.
50%UnverifiedAcross evaluations in three domains—robotic manipulation, autonomous navigation, and text-to-video generation—EdgeWise delegated 10-20% of steps to the cloud, maintaining over 95% of original performance with roughly an 85% cost reduction.
50%UnverifiedEdgeWise reported a 95th percentile latency of 190ms for edge-side steps, a cloud call peak of 180ms, and end-to-end latency below 200ms for 94% of steps.
50%UnverifiedEdgeWise's risks include selector overfitting to benchmark domains, OS and hardware fragmentation, unpredictable cloud cost spikes, the emergence of all-edge LLMs, and security and data privacy compliance.
50%UnverifiedEdgeWise is an edge-first long-horizon Agent solution that uses a step selector to determine whether the current reasoning step needs a cloud large model, thereby compressing cloud calls to 10-20% of total steps.
50%UnverifiedUnder the traditional all-cloud architecture, each picking action consumes 1200 tokens, costing $0.0096; after EdgeWise selects delegation, each round consumes 300 tokens, costing $0.0012.
50%UnverifiedThe EdgeWise selector is evaluated using ranking metrics such as MRR (Mean Reciprocal Rank); when MRR approaches 1, it can precisely route key steps to the cloud.
50%UnverifiedEdgeWise chooses the open-source ONNX Runtime inference engine, letting partners compile the selector themselves on platforms such as Linux, RT-OS, iOS, and custom ASICs.
50%UnverifiedEdgeWise provides an Apache 2.0-licensed 12M-parameter pretrained selector hosted on HuggingFace, as well as reference Docker images for edge and cloud.
50%UnverifiedThe EdgeWise roadmap shows: public Beta in Q4 2026, SDK v1.0 release in Q1 2027, and a broader cloud orchestration service in Q2 2027.
50%