Breaking

DeepReinforce Releases Self-Improving Coding LLM Family Ornith-1.0 Under MIT

June 25, 2026 at 10:33 EDT

Open Source
Software Dev & Coding
Foundation Models

DeepReinforce on June 25, 2026 released Ornith-1.0, a self-improving family of open-source LLMs designed for agentic coding. Four MoE/Dense models ranging from 9B to 397B were all released under the MIT license, allowing free commercial and research use.

June 25, 2026 · DeepReinforce

Ornith-1.0: Self-Improving Open Models That Write Their Own Scaffolds

A family of four MIT-licensed agentic-coding LLMs (9B–397B). The flagship 397B MoE claims coding performance on par with Claude Opus 4.7 — trained with RL to generate task scaffolds itself, not rely on human-designed harnesses.

models, 9B to 397B

397B

flagship MoE, Opus-4.7-class coding

MIT

fully open weights, free commercial use

SWE-Bench Verified — scaling with size

Score (%) climbs as model size grows across the lineup

69.4

9B Dense

75.6

35B MoE

82.4

397B MoE

Terminal-Bench 2.1 — same trend

Score (%); the 35B MoE even surpasses the larger Qwen 3.5-397B here

43.1

9B Dense

64.2

35B MoE

77.5

397B MoE

The self-improving loop

RL jointly optimizes the scaffold and the solution — the model learns how to solve a task, not just the answer.

Generate scaffold

prompts · workflows · tools

→

Solution rollouts

solve the task

→

RL reward

refine both jointly →

WHAT DEVELOPERS LIKE

Self-improving RL approach to scaffolding
MIT license — easy commercial use & fine-tuning
Open weights, GGUF builds for local inference

STILL UNPROVEN

Release is very recent — no long-term track record
Real-world SWE-Bench usage yet to be assessed
Scaffold quality & limitations await user feedback

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.