The landscape of open-source language models is evolving at breakneck speed. With the arrival of Ornith-1.5-35B-A3B, a new milestone has been set in the quest for models capable of continuous self-improvement. This model redefines the standards for AI-assisted software development and agentic capabilities.\n\n## What is Ornith-1.5?\n\nOrnith-1.5 represents a major evolution. The key innovation lies in the joint optimization of three pillars: task generation, scaffolding construction, and solution rollouts via reinforcement learning (RL).\n\n## Performance: The bet on efficiency\n\nThe Ornith-1.5-35B-A3B model is a Mixture-of-Experts (MoE) type model that activates only about 3 billion parameters per token. Despite this lightness, it shows impressive results, often outperforming much larger models on coding benchmarks such as SWE-bench Verified (79.0) and GPQA Diamond (89.2).\n\n## Why is it so efficient for development?\n\nThe strength of Ornith-1.5 lies in its architecture, which is designed for reasoning. Unlike classic models, it natively integrates a thought process (Chain-of-Thought) identifiable by <think> tags. This capacity is crucial for navigating massive codebases or resolving complex bugs.\n\n## Technical implementation\n\nTo leverage Ornith-1.5, it is necessary to have an up-to-date environment (Transformers 5.8.1, vLLM 0.19.1). The model natively supports a context window of 262,144 tokens, which is ideal for analyzing full code repositories.\n\n## Conclusion\n\nOrnith-1.5 confirms that performance gains no longer come solely from the number of parameters, but from data quality and the model's ability to self-correct. It establishes itself as an essential tool for developers wishing to integrate powerful agents into their pipelines.
Ornith-1.5-35B-A3B: The Self-Improving AI for Coding
Discover Ornith-1.5-35B-A3B, the new open-source model redefining self-improvement and agentic coding performance, surpassing models of equivalent size.