Ornith 1.5 Explained: The Self-Improving Open Model

Описание: Ornith gave away the weights to a model family that trains itself, and in this beginner's guide I walk through what Ornith 1.5 actually is, how the self-improvement loop works, and the four traps nobody puts in the announcement. We cover the three sizes — a 9B dense model that fits on a laptop, a 35B-A3B mixture-of-experts that activates about 3B parameters per token, and a 397B MoE flagship, plus a quantized 9B Mobile build for phones — then the core mechanism, where the model proposes its own task, generates a task-specific scaffold, produces solution rollouts, and updates the policy with GRPO, with task rewards combining validity, frontier difficulty targeting roughly a 20 percent success rate, and novelty. We also cover the self-reported benchmark table, the 262,144-token native context and the YaRN factor 4.0 extension to about 1M tokens, running it locally with Ollama and llama.cpp GGUF quants, serving it properly with vLLM or SGLang including the tool-call parser and reasoning parser flags that beginners always miss, the two published sampling profiles for general use versus precise coding, and the think block that arrives as reasoning_content. Everything here is grounded in the official Ornith blog post, the Hugging Face model cards and the Ollama library page — links below. Commands, sampling settings and chapters are in this description. Keywords: ornith 1.5, ornith ai, open weights, open source llm, mit license model, self improving ai, self scaffolding, grpo, reinforcement learning, mixture of experts, moe, 9b model, 35b a3b, 397b, local llm, ollama, llama.cpp, gguf, quantization, vllm, sglang, tool calling, reasoning model, agentic coding, terminal bench, swe-bench verified, gpqa diamond, long context, yarn rope scaling, run llm locally, beginners guide, tips tricks traps. If this helped, hit like and subscribe.