This page may contain stale information. Last updated: 2026-06-27

Overview

Ornith-1.0 is an open-source coding model family from deepreinforce, released June 25, 2026 under MIT license on Hugging Face. Built on gemma-4 and qwen 3.5 foundations.

Model Sizes

  • 9B Dense — edge deployment
  • 31B Dense
  • 35B MoE
  • 397B MoE — flagship

Self-Scaffolding RL

Instead of fixed human-designed agent harnesses, Ornith learns to generate task-specific scaffolds during RL. Two-stage RL step: (1) refine scaffold; (2) generate solution. Reward propagates to both stages.

Benchmark scores (82.4 SWE-Bench Verified, 77.5 Terminal-Bench 2.1 for 397B) are vendor-reported. Independent reproduction pending. Trails Claude Opus 4.8 on several benchmarks.

Anti-Reward-Hacking

  1. Fixed trust boundary (immutable outer layer)
  2. Deterministic monitor (zero reward for reading withheld tests)
  3. Frozen LLM judge as veto

Sources