Ornith-1.0 Agentic Coding LLM Evaluation: 35B MoE Superiority
Generated: 2026-06-29 · API: Gemini 2.5 Flash · Modes: Summary
Ornith-1.0 Agentic Coding LLM Evaluation: 35B MoE Superiority
Clip title: Ornith 1.0 First Look & Test – The BEST New Local Coding Models? Author / channel: Bijan Bowen URL: https://www.youtube.com/watch?v=Y7VL_whmAqQ
Summary
The video provides an extensive first look and testing session for the Ornith-1.0 family of open-source Large Language Models (LLMs), specifically designed for agentic coding. The presenter focuses on evaluating two available variants: the 9-billion parameter (9B) dense model and the 35-billion parameter Mixture-of-Experts (MoE) model. The context for this evaluation is the increasing importance of accessible open-weight models, given growing restrictions on proprietary frontier intelligence. Ornith-1.0 distinguishes itself by employing a “self-improving training framework” that jointly learns to solve tasks and construct optimal solution paths, utilizing Reinforcement Learning with Group Relative Policy Optimization (GRPO).
Across several demanding agentic coding tasks, the 35B MoE model consistently demonstrated superior capabilities compared to its 9B dense counterpart. For instance, in the “Browser OS Test,” the 35B MoE generated a much more complete and aesthetically pleasing WebOS, featuring functional applications like a notepad, a web browser that opened to Wikipedia, a wallpaper changer, a file manager, and an interactive desktop pet with trackable stats. While its “GTA clone” was 2D instead of 3D, it showcased interesting immersive elements and character animations, alongside a functional 3D dungeon crawler. Similarly, for the “Slappis Watch Co. Website” task, the 35B MoE produced an impressive, visually appealing website with a spinning 3D watch model and correctly rendered product cards, a feat that had previously challenged even larger frontier models.
In contrast, the 9B dense model struggled significantly with many of the complex coding challenges. Its attempt at the “Browser OS Test” resulted in a visually “oddly looking” interface with many non-functional elements. It faced repeated errors and often entered “thought loops” when attempting to debug in LM Studio. For tasks like creating an FPS game from a subway scene or generating a 3D model from an image, the 9B model yielded either completely broken, unplayable results or very rudimentary, incomplete models. While the OpenCode agentic debugger showed some promise in fixing issues generated by both models on certain tasks, it couldn’t consistently salvage the 9B’s more fundamentally flawed outputs.
In conclusion, the 35B MoE variant of Ornith-1.0 emerges as a promising open-weight model for agentic coding, capable of generating more complex and functional web-based applications and 3D scenes than anticipated for its size, especially when run unquantized. The video highlights a genuine improvement in its performance over the baseline Qwen models it is fine-tuned from. Although both models still exhibit limitations in achieving perfect execution of complex requests without errors or requiring debugging, the 35B MoE’s ability to produce highly decent results for advanced front-end and 3D rendering tasks is notable. The presenter expressed keen interest in future, potentially more potent versions, such as the unreleased 31B variant based on Gemma-4, which could further push the boundaries of accessible open-source agentic capabilities.
Video Description & Links
Description
Timestamps:
00:00 - Intro 00:53 - Note on AI Situation 02:13 - First Look 03:00 - Technical Look 04:51 - Browser OS Tests 13:01 - Subway Scene Tests 15:55 - Subway FPS Tests 16:35 - 35B Subway FPS Fix 18:22 - Watch Frontend Tests 20:40 - 35B C++ Test 22:10 - 9B Multimodal Code Test 22:55 - 9B Agentic Coding Tests 26:53 - Results Overview 29:37 - Closing Thoughts
AI Integration & Consulting: https://bijanbowen.com/ Join the Discord: https://discord.gg/hfaR2exy7S
In this video, we take a hands-on look at Ornith 1.0, testing the new local coding-focused models across a range of practical tasks.
We begin with a first look and technical overview, then move into browser-based workflows, subway scene generation, FPS-style game tests, frontend design tasks, C++ coding, multimodal code understanding, and agentic coding scenarios.