Move AI execution state across models — without starting over.
When a workflow hands off from one model to another, the target normally re-reads the entire task history from scratch. UnimAI Relay carries the reusable computed state across the boundary and recomputes only the part that has to change.
faster handoffs
138 ms vs 899 ms native Qwen 7B prefill
quality retained
43.97% vs 45.69% native accuracy
One task history, rebuilt at every model boundary
Multi-agent systems route planning, retrieval, coding, and verification to whichever model fits each step best. That routing buys capability and cost — then gives it back, because each handoff reprocesses the same long history from scratch.
Relay carries the reusable computed state across the boundary, verifies it against a quality threshold, and recomputes only the delta that threshold demands.
Existing caching systems reuse within a model.
UnimAI reuses across them.
- What today's systems do
- Prefix and KV caching in engines like vLLM, SGLang, and LMCache reuse computed state when the same model sees the same prefix again. Reuse is a lookup inside one model's representation space.
- Where they stop
- A handoff to a different model has no shared representation — different tokenizer, different attention geometry, different hidden dimensions. There is nothing to look up, so the target model performs a full prefill of the whole history.
- What Relay adds
- Relay maps state between representation spaces, checks the result against a quality threshold, and recomputes only the delta that check demands — falling back to full prefill when a model pair cannot clear the bar.
Where it pays off
Model routing
Send each step to the best-fit model without paying a full prefill at every switch.
Agent handoff
Planner, coder, and verifier share one task history instead of rebuilding it three times.
Cost-aware serving
Run more of the workload on smaller targets while retaining most of the larger model's quality.
Built by systems and agent researchers
Four founders spanning AI infrastructure, distributed systems, agentic AI, and enterprise execution — with work published at ICLR, MLSys and ACL.
Dr. Yi Zhao
CEO & Co-founder
Dr. Bingzhe Li
CTO & Co-founder
Yi Li
Co-founder
Dongming Jiang
Co-founder
Work with us
Relay is a working prototype with measured cross-family results. The next step is proving it on real workloads.
Design partners
Teams running context-heavy, multi-model production traffic:
- multi-model inference
- agent infrastructure
- heterogeneous serving systems
Investors
We're a pre-seed/seed team and happy to share:
- the full benchmark methodology and per-pair results
- a technical deep-dive on cross-family translation
- roadmap and commercial plan
Either way, use the form or email info@unimai.ai.