EDGE INTELLIGENCE · CLOUD–EDGE COLLABORATION

Edge Intelligence

边缘智能

Build a software–hardware edge collaborative inference framework that breaks the latency–efficiency–privacy trilemma for large-model deployment.

Core Challenges

Bringing large models to devices and industrial scenes requires answers on efficiency, latency, compute, and security.

Efficiency

How can devices use large-model capability while reducing token consumption and improving token efficiency?

Latency

How can we effectively reduce time-to-first-token (TTFT)?

Compute

How can we improve the inference speedup ratio?

Security

How can we better protect local private data?

Solution & Architecture

Heterogeneous dual-brain on device with cloud LLMs: local draft prediction and cloud parallel verification form an efficient, secure, low-latency loop.

Edge · Dual Brain

On-device lightweight model (e.g. Qwen-0.8B) runs draft prediction

Cloud · Large Model

Cloud LLM verifies token sequences in parallel

Collaborative Inference Workflow

01
Local Data Processing
Desensitization / compression / feature extraction on the AI inference chip
02
Draft Prediction
Generate token sequences locally
03
Cloud Verification
Token sequences are verified in the cloud
04
Return Results
Efficient, secure, and low-latency

Core Technologies

From algorithms to chips—building deployable edge collaborative inference.

Core Technologies

Heterogeneous Dual-Brain & Speculative Sampling

Turn serial inference into local parallelism: draft on device, verify in cloud, raise edge–cloud throughput.

Core Technologies

Semantic Compression & Token Efficiency

Convert raw data into high-density desensitized embeddings to cut wasteful token cost.

Core Technologies

Privacy Barrier & Feature Mapping

Hardware-level isolated encrypted compute—“data stays on device; only features go to the cloud.”

Core Technologies

On-Device AI Inference Chip Architecture

Hardware optimized for Transformer ops, with efficient KV-cache management and low-power matrix compute.

Expected Outcomes & Value

Target metrics that describe efficiency, experience, and trust gains from edge collaboration.

2~3x

Target inference speedup

TTFT ↓

Significantly lower TTFT

50%+

Projected cloud token savings

Highest Tier

Local-first privacy (data on device; features only upstream)

Loading…