A small team just halved hallucination rates on a 27B parameter model based on Qwen, trained entirely on open-source datasets. It outperforms GPT-6 Luna on Terminal-Bench while generating 50% fewer tokens per task—a massive efficiency win for local inference.
Their v2 architecture runs 7 concurrent long-context agent sessions on a single consumer GPU. Full Apache 2.0 license.
This is the local AI endgame: smaller teams shipping production-grade models that actually run on your hardware without API costs or vendor lock-in. Open weights, open training stack, zero compromise on performance.
Their v2 architecture runs 7 concurrent long-context agent sessions on a single consumer GPU. Full Apache 2.0 license.
This is the local AI endgame: smaller teams shipping production-grade models that actually run on your hardware without API costs or vendor lock-in. Open weights, open training stack, zero compromise on performance.