FuturixAI has released CARVE, an open-source model family designed for structured inference, built upon Qwen3.8-27B and licensed under Apache 2.0. This model family aims to generate structured JSON outputs that can be parsed into typed data, offering an alternative to approaches like Jev that directly score options. CARVE achieved 82.22% accuracy on the Decision 1.0 transfer-v9 benchmark, with reported latencies of 72.5 ms p50 and 91.0 ms p95 on specific NVIDIA hardware. AI
IMPACT Provides an open-source alternative for structured AI inference, potentially enabling new applications and research.
RANK_REASON Release of an open-source model family with benchmark results. [lever_c_demoted from research: ic=1 ai=1.0]
- Apache Software License 2.0
- Carve
- Choice
- Decision 1.0
- FuturixAI
- Jevíčko
- Kev-9B
- Noulens
- NVIDIA RTX PRO 6000 Blackwell Server Edition
- Qwen3.8-27B
- System One API
- TypeSafe AI
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →