Show HN: Open-source model routing for coding agents at Astra-level performance
Weave Router 2.0 is an open-source dynamic model router for coding agents that matches frontier model performance at half the cost and double the speed.
summary is enough It provides a clear overview of the routing architecture and benchmarks, though deep implementation weights remain proprietary.
summary
Weave Router 2.0 is an open-source tool that plugs into coding agents to dynamically switch between LLMs based on task complexity, cost, and cache state. Benchmarked against GPT-6 Astra, the router achieves equivalent pass rates on Terminal Bench 4.0 and SWE Atlas while cutting costs in half and running over twice as fast. The core architectural innovation uses a hidden Markov model to track session states and narrow down model buckets, significantly shrinking the routing search space. Additionally, a specialized cache-eviction impact calculator helps prevent unnecessary model switches that would waste context cache.
key points
- Weave Router 2.0 matches GPT-6 Astra pass rates on Terminal Bench 4.0 and SWE Atlas.
- It achieves these benchmarks at 52-54% of the cost and 2.2x to 2.5x faster execution speed.
- A hidden Markov model (HMM) tracks session state to drastically reduce the combinatorial search space for routing decisions.
- A cache-eviction impact subsystem accurately calculates the cost of switching models to preserve context caches.
who should read it
Engineers and developers building AI coding agents or looking to optimize LLM API costs and execution speed.
discussion positive
consensus
Commenters are highly interested in the architecture and compare it to existing solutions like Cursor and Copilot auto modes, focusing heavily on cost optimization and provider variance.
pushback
Some question how a router trained on frontier model labels can ultimately exceed frontier performance, or note limitations of per-session routing.
If you are training on data labeled by frontier models, how do you expect to exceed the performance of frontier models, other than in the cost dimension by recognizing simpler problems and routing to cheaper models?
It raises a fundamental question about whether a routing model trained on frontier model outputs can ever surpass the underlying models it orchestrates.
Generated by an LLM from the article and thread. Check the sources before relying on it.