Education

New Agentic Coding Model Ornith 9B — Is It Worth Running Locally?

by Bart Slodyczka

📚 Main topics

Model and benchmark overviewThe video examines Ornith 1.0, an open-weight model family designed for agentic coding, focusing on the 9B model. The presenter questions how much published benchmarks reveal about building something from scratch and tests the model on a real coding task instead. 0:00
Local setup and performanceThe 9B model is run through LM Studio on an M4 Mac Mini with 16 GB of RAM, using about 12 GB for the model. The presenter reports roughly 16 tokens per second and says a long context window is possible on this hardware. 3:39
Tower defense coding testUsing Pi agent, the presenter asks the 9B model to create a single-file HTML tower defense game, then compares its work with the 35B model running on a more powerful Mac Studio. 7:17

✨ Key takeaways

The 9B model struggles to deliver a working gameIts initial HTML file looked comparable in size to the 35B model’s, but the game did not function correctly. The 35B version worked after a fix to the towers’ attacks. 11:24
Debugging was difficultThe presenter found that the 9B model struggled to repair its own code and could get stuck repeating attempts. The 35B model was used over several iterations to salvage the game. 13:00
Bigger models were more capable and faster in this comparisonThe 35B model generated at about 100 tokens per second on the more powerful machine, compared with around 16 tokens per second for the 9B model on the Mac Mini. 10:22

🧠 Lessons learned

Small models need narrower tasksThe 9B model may be useful for coding, but the presenter found that a complex task requiring many pieces to work together exceeded its accuracy and reliability. 14:34
Watch for implementation mistakesThe presenter observed problems such as misspelled or missing functions, empty functions, and functions that were defined but not used. These issues can make tool-based coding workflows harder to debug. 13:00

🏁 Conclusion/next steps

Set expectations according to model sizeFor the tower defense task, the 9B model did not produce a working result on its own, while the 35B model showed stronger coding ability. The presenter’s takeaway is to reduce expectations or the scope of work when using a 9B model locally. 15:04

Transcript excerpt

0:00 So, there's an interesting new open- source model called Ornith 1.0, which is meant to be really good at agentic coding. So, in this video, we'll be testing out the 9 billion parameter dense model on a 16 gig Mac Mini to see just how good it is. So, Ornith 1.0 is a brand new family of open source models, which is specialized for agentic coding. They were released by a company called Deep Reinforce, which is based in San Francisco. And Ornith comes out in four different sizes. The 9 billion parameter dense, which is what we'll be testing today on a 16 gig Mac Mini. We also have the 31B dense, 35BOE,

0:32 and 397BOE. I'm pretty sure these three, the 9B, 35B, and 397B are open weight, which means you can just download them and run them today on your own device. And just the 31B is not released, maybe not released yet, and it might be released later in the future. So, there's some pretty interesting claims throughout this release paper. Uh, the first being that Ornith 1.0 was built on top of a pre-trained Gemma 4 and Qwen 3.5. Gemma 4 in my opinion is one of the best agentic task completion models. So you plug it into your Hermes agent or

🔒 The full, searchable transcript is available with Pro.

Suggestions