GitHub radar
Ornith-1.5-122B: an AI that invents its own tasks and learns
The anonymous team ornith-ai released a model trained through full self-improvement: it sets its own tasks and learns to solve them instead of copying human examples. After this guide you'll know if it fits your work and how to try it if you have the right hardware.
It's a large language model (a neural network for text and code) with 122 billion parameters that invents its own exercises and learns to solve them through reinforcement learning — reward for correct answers.
When it helps
Useful if you're curious about experimental AI self-learning and have a server-grade 80GB GPU. Not for real work — the results aren't independently verified, and the team has no release history.
Pros
- Claims to beat well-known models on coding and reasoning benchmarks
- Can call external tools like a real AI agent
- Handles context up to 262,144 tokens (word pieces)
- Runs on a single GPU instead of a server cluster
Cons
- Benchmark results come only from the team itself, no independent checks
- ornith-ai has no prior releases to judge their track record
- Needs a server-grade 80GB GPU — a regular laptop won't run it
How to set it up — step by step
- 1Open the model page from the facts link and check the Quickstart section.
- 2Check whether you have a server with an 80GB GPU — without it the model won't run.
- 3Send the agent the text from the block below to learn about its self-learning method and get an honest answer about running it without a powerful GPU.
- 4If you lack that hardware, treat the model only as inspiration about AI self-learning, not as a working tool.
- 5Before trusting the benchmarks, look for independent reviews — ornith-ai has no release history yet.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Look into this model for me: https://huggingface.co/ornith-ai/Ornith-1.5-122B — explain what its self-improvement training method actually does, and tell me honestly whether it can run without a server-grade 80GB GPU.
This is a server-grade model, not something you can run at home. It needs a single GPU with 80 GB of memory.
Open on Hugging Face▌ More finds