AI Model2026
A text-only open-weight MOE code model (35B total, 3B active) fine-tuned with SFT+RL for agentic coding; achieves strong agentic-code benchmarks, supports 262k context and deployment via Transformers/vLLM; vision weights are not included.