Wait, This model distilled 4.6 Opus into an Open 9B Model & it works with Claude Code?!
Omnioder 9B is a new open-source coding agent model built on QN359B that was fine-tuned on 425,000+ curated coding trajectories from frontier models like Claude Opus 4.6 and GPT models. While it shows promising improvements over its base model and offers Apache 2.0 licensing with local deployment options, the creator advises skepticism about self-reported benchmarks and notes inconsistencies in the reported numbers.
Summary
The video analyzes Omnioder 9B, a new open-source coding agent model released on HuggingFace by Tesla. Unlike typical code completion models, this is positioned as a true coding agent that was fine-tuned on over 425,000 curated agentic coding trajectories from successful runs of frontier models including Claude Opus 4.6, GPT 5.4, GPT 5.3, Codeex, and Gemini 3.1 Pro across various scaffolds like Claude Code and Open Code. The model is built on QN359B with a hybrid architecture featuring gated delta networks and standard attention, claiming a native context length of 262,000 tokens. It supports thinking mode with think tags and is Apache 2.0 licensed with GGUF files available for local deployment. The key value proposition is that the model learned disciplined coding agent behaviors like 'read before write' patterns, responding to LSP diagnostics, and making minimal diffs rather than rewriting entire files. Benchmark results show improvements over the base model, with 83.8% on GPQA diamond and 23.6% on TerminalBench 2.0, though it still trails some other models like GLM 4.7 and Claude Haiku 4.5. The creator notes inconsistencies in reported benchmark numbers on the same page and emphasizes the need for independent testing. The model has limitations including non-English performance concerns and works best with scaffolding patterns similar to its training data. The creator views this as a promising release for local deployment and experimentation, particularly for users who want an open alternative to subscription-based coding tools.
Key Insights
- Tesla trained Omnioder 9B on over 425,000 curated agentic coding trajectories from frontier models like Claude Opus 4.6, GPT 5.4, and others, attempting to compress frontier coding agent behavior into a 9 billion parameter open model
- The model learned disciplined coding agent patterns like 'read before write' behavior, responding to LSP diagnostics, and preferring minimal diffs instead of rewriting everything, which addresses common problems with weaker models that panic and rewrite whole files
- There are inconsistent benchmark numbers reported on the same HuggingFace page, with TerminalBench 2.0 showing 23.6% in the main section but 28.1% in the evaluation metadata, which raises questions about the reliability of self-reported benchmarks
Topics
Full transcript available for MurmurCast members
Sign Up to Access