Summary
Fahd Mirza runs a direct head-to-head evaluation between Claude Sonnet 5 and Ornith 1 35B — a new open-source mixture-of-experts coding model from Deep Reinforce — to answer a practical question: how close can a locally-hosted open-source model get to one of the best closed-source agentic models currently available?
The first test is an autonomous debugging challenge using a live full-stack call center application. Mirza plants four hidden bugs: a misspelled column name in the database query (line 70), a wrong HTTP method on a route handler (GET instead of POST, line 85), a division error in an analytics calculation (line 146), and a port misconfiguration pointing to 9000 instead of 8000. He also plants a deliberate red herring — a misleading code comment falsely claiming that transferred calls are double-counted — to test whether each model trusts documentation or verifies the underlying logic. Both Claude Sonnet 5 and Ornith 35B identify and patch all four bugs correctly, and neither model falls for the misleading comment.
The second test is single-shot code generation: build a working animated train junction simulation with trains from Mumbai and Germany, spinning wheels, track layouts, and smooth motion — all in one self-contained file. The video concludes that for the majority of coding tasks, a free, private, locally-hosted model running on commodity hardware can match frontier closed-source performance, with only minor differences in agentic loop hygiene. For developers weighing cost, data privacy, or API dependency, the results make a meaningful case for local alternatives.
📺 Source: Fahd Mirza · Published July 01, 2026
🏷️ Format: Comparison
