OmniCoder-9B brings agentic coding to 8GB GPUs

A 9B model trained on GPT-5 and Claude Opus 4.6 agentic traces that fits in 8GB of VRAM is exactly what the self-hosted coding movement has been waiting for.

• –Q4_K_M quantization lands at ~5.7GB, running comfortably on RTX 3070/4060-class cards — no cloud API required
• –Terminal-Bench 2.0 score of 23.6% is a 61% improvement over the Qwen3.5-9B base (14.6%), suggesting real agentic gains rather than benchmark overfitting
• –Training on scaffold patterns from Claude Code, OpenCode, and Codex effectively distills frontier agentic behavior into a local-first model
• –Native 262K context window (extensible to 1M+) is exceptional at this size class and critical for multi-file coding sessions
• –Apache 2.0 license with OpenAI-compatible llama-server API means drop-in replacement for existing Cline/VS Code setups with zero vendor lock-in