Ornith-1.5: Open Models for Coding and Agents

The Ornith team released the Ornith-1.5 family of open models for coding and agents on August 18, 2026, in three versions (9B, MoE 35B-A3B, MoE 397B), trained in a self-improvement loop using GRPO; Ornith-1.5-397B scores 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, with weights published on Hugging Face.

Compiled by Sergey KostenchukPublished 2026-08-20Updated 2026-09-18

2026-08-20 Research Anthropic

🤖 Ornith-1.5: Open Models for Coding and Agents

The Ornith team released the second generation of open Ornith models for coding, logic, and agents on August 18, 2026. Three versions: dense 9B, MoE 35B-A3B (3B active parameters per token), and MoE 397B, with 128K-256K context. The key mechanic is a self-improvement loop: the model itself generates tasks, writes scaffolds for them, and trains via GRPO on its own attempts.

🌍 According to the developers, Ornith-1.5-397B scores 86.1 on Terminal-Bench 2.1 — above the level of Claude Opus 4.8 (85.0) — and 56.0 on DeepSWE. Weights are open in FP8, GGUF, MLX, and NVFP4 formats.

👤 The 9B version in quantization runs on a smartphone, 35B-A3B outperforms Gemma 4-31B on SWE-Bench Verified (79.0 vs. 52.0), and 397B is designed for local servers. Everything is already available for download on Hugging Face.

Source 1: https://ornith.ai/ornith_1_5.html Source 2: https://huggingface.co/collections/ornith-ai/ornith-15

Sources