OmniCoder-9B | 9B coding agent fine-tuned on 425K agentic trajectories

Tesslate has released OmniCoder-9B, a specialized 9-billion parameter coding agent fine-tuned on the Qwen3.5-9B architecture. What sets this model apart is its training data: 425,000+ curated agentic coding trajectories collected from real-world software engineering tasks performed by frontier models including Claude Opus 4.6, GPT-5.4, GPT-5.3-Codex, and Gemini 3.1 Pro. These trajectories capture complex behaviors like tool use, terminal operations, and multi-step reasoning across platforms like Claude Code, OpenCode, and Droid.

The model inherits Qwen3.5's hybrid architecture with Gated Delta Networks interleaved with standard attention, enabling efficient processing of its full 262,144 token context window (extensible to 1M+). More importantly, OmniCoder-9B demonstrates sophisticated agentic behaviors learned directly from its training data: it implements read-before-write patterns to recover from errors, responds to Language Server Protocol (LSP) diagnostics, and applies minimal edit diffs instead of full file rewrites. It also supports structured reasoning through <think>...</think> chains for complex problem decomposition.

Available under the permissive Apache 2.0 license on Hugging Face, OmniCoder-9B represents a significant step toward more reliable, production-ready AI coding assistants. By learning from successful agent trajectories rather than just code completion, it bridges the gap between traditional code generation models and fully autonomous coding agents capable of handling real software development workflows.