A new local Ai model

@hatsunama

3.2K followers · usually 1.1K views a post

Views

3.3K

Likes

205

Comments

33

Shares

49

SoundTikTok sound · original sound - hatsunama

Promoted product

Ornith 1.0

Complete Ornith 1.0 guide: self-scaffolding AI models for agentic coding. 9B to 397B sizes, MIT licensed, benchmark results, local setup, and model comparison.

3 videos3 creators

Why it worked

The video highlights a novel and advanced feature of the AI model (self-improving scaffolding) which is presented as a significant technical achievement, appealing to an audience interested in cutting-edge AI developments. The use of on-screen text and clear explanations makes complex information accessible.

Summary

The video introduces Ornith-1.0, a family of open-source AI models for coding that can be run locally. It highlights the models' performance, particularly the flagship 378B parameter version, and emphasizes its unique self-improving training framework where the AI designs its own scaffolding.

Structure

1Introduction of Ornith-1.0 as a local AI model
2Comparison of model sizes and performance benchmarks
3Explanation of the self-improving training framework

Original caption

A new local LLM for your home computers! Get em before the US government says they're evil 😆 #ai #news #local #llm #fypシ

More from @hatsunama

Product placement

The video prominently features the Ornith-1.0 AI model, showing its name, logo, and performance benchmarks on screen. The speaker discusses its features and capabilities.

Call to action

Hit that save button. Let's get in.

On-screen text

Ornith-1.0 An Open-Source Coding Model Family That Learns Its Own RL Scaffolds RELEASED v1.0 LLM Performance Evaluation Grade 1-10B GPT-4 Gemini-0.5B Gemini-1.5B Gemini-Pro Gemini-Ultra Terminal Bench 2.1 (Sparrow Z) SWE-Bench (Verified) SWE-Bench (Pro) SWE-Bench (Multilingual) NL2Repo Commonsense Avg SWE Atlas - QnA SWE Atlas - TW Highlights: State-of-the-Art Coding Agents: Available in 9B-Dense, 31B-Dense, 35B-MoE, and 397B-MoE (post-trained on top of Gemma 4 and Qwen 3.5), achieving state-of-the-art performance among open-source models of comparable size on coding benchmarks such as Terminal-Bench 2.1, SWE-Bench, NL2Repo and OpenClaw. Self-Improving Training Framework: Ornith- 1.0 employs RL to learn to generate not only solution rollouts, but also the scaffold that drive those rollouts. By jointly optimizing the scaffold and the resulting solution, the model discovers better search trajectories and generates higher-quality solutions. Things like memory management tool use and calling logic error handling and retry strategies workflow orchestration and even prompt structuring Traditionally engineers would have to design these by hand. But with Ornith 1.0 it actually designs and redefines and redesigns these over and over as it goes through the training process. So what automatically emerges from this is better debug flows, smarter retry logic, optimized tool sequences, etcce The scaffolding co evolves with the Model's reasoning

Transcript

A new local Ai model just dropped and you can actually use this one at home. It's beating all the benchmarks possible. Hit that save button. Let's get in. Today we're talking about Ornith-1.0 and this is an AI model that you can run at home right now. It's actually not just an AI model, it's an entire family of models. They have a tiny model that's only 9 billion parameters and then a massive beast of 378 billion parameters and everything in between. As far as benchmarks, the flagship 378 billion parameter model does do better than Opus 4.7 and the smaller versions do really well to punch far above their weight. But that's not the craziest part. The best part is that this thing learns to create its own RL scaffolding during training. What that means is the orchestration framework or the wrapper around the AI that helps it complete long-term complex tasks is being built by the AI as it's learning the things that it's learning during training. Things like memory management, tool use and calling logic, error handling and retry strategies, workflow orchestration and even prompt structuring. Traditionally engineers would have to design these by hand. But with Ornith 1.0 it actually designs and redefines and redesigns these over and over as it goes through the training process. So what automatically emerges from this is better debug flows, smarter retry logic, optimized tool sequences, etc. The scaffolding co-evolves with the model's reasoning. Mind is blown.