LinkAnnouncement
Instead of relying on fixed, human-engineered harnesses across coding tasks, treating the scaffolding as a co-evolving learnable object allows agentic workflows to autonomously discover more effective execution strategies.
The benchmark data demonstrates strong scaling characteristics across all sizes: the 9B variant matches or exceeds previous 31B models, and the 35B MoE outperforms larger architectures like Qwen 3.5-397B on Terminal-Bench 2.1. By restricting learnable scaffolding strictly to inner orchestration and isolating test boundaries, the training pipeline avoids common reward-hacking failure modes.
Summary generated by AI (Gemini 3.8 Flash, 27 September 2026). It may contain errors – the original source is authoritative.
Similar in content, found by the AI search.
Link:ornith-ai
ornith-ai has released Ornith-1.5-397B, an open-source multimodal mixture-of-experts model under the MIT licence. With 397 billion parameters and an end-to-end self-improvement loop, it delivers coding and reasoning benchmark performance comparable to top proprietary models.
Repository:ornith-ai/Ornith-1
The Ornith-1 repository provides an open-source model family designed for agentic tasks using autonomous self-improvement. Spanning Ornith-1.0 and Ornith-1.5, the project implements self-scaffolding and automated generation of training rollouts for reinforcement learning.
2kstars
Article:Luis Chavez-Mattos
Deep Reinforce has introduced Ornith 1.5 35B-A3B, a mixture-of-experts model activating roughly 3 billion parameters per token out of 35 billion. According to published benchmarks, it surpasses direct competitors like Qwen3.6-35B across coding and agentic tasks, while closing the gap on much larger models such as Qwen3.5-397B.
Link:huggingface.co
Ornith-1.0 ist eine Sammlung von Open-Source-Sprachmodellen, die speziell für agentenbasierte Programmieraufgaben entwickelt wurden. Die Kollektion umfasst verschiedene Varianten, darunter das Modell Ornith 1.0 9B, das für logikbasierte Unterstützung bei der Softwareentwicklung optimiert ist.
Link:Kwaipilot
Kwaipilot has released KAT-Coder-V2.5-Dev, an open-weight Mixture of Experts model based on Qwen3.6-35B-A3B. Featuring 35 billion total and 3 billion activated parameters, it has been specifically post-trained for agentic software engineering tasks.
Article:Seth Karten, Alex L. Zhang, Kevin Thomas, Sebastian Müller, Prime Intellect Team
Prime Intellect has released Prime Agent, an open-source coding harness built on two abstractions: Recursive Language Models (RLM) and Continual Harness. Instead of relying on rigid tool-calling schemas, the agent operates programmatically inside a persistent IPython kernel and dynamically adapts its own configuration during execution.
Parts of this content were created with the assistance of AI.
