AI Tools Review

Ornith 1.5

Version: 1.5

By DeepReinforce

Released: 2026-08-19

Open Weights
MIT Licence
Self-Improving
Local AI
Coding
Free
New

DeepReinforce's MIT-licensed, self-improving open-weights model family, released 19 August 2026 in three sizes (9B dense, 35B MoE, 397B MoE flagship). Its training loop has the model generate its own tasks, scaffolds and reinforcement-learning rollouts, and it posts benchmark scores that match or beat Claude Opus 4.8 on some agentic and coding tasks while trailing it on others.

Visit Ornith 1.5

AI-Powered

Leverages advanced AI technology to deliver cutting-edge capabilities and results.

Fast & Efficient

Optimized performance ensures quick results without compromising on quality.

Purpose-Built

Specifically designed for llms tasks and workflows.

DeepReinforce Model Timeline

Ornith 1.5Current

Specifications

pricingFree, MIT licensed; self-hosting compute cost scales with model size (9B up to 397B)

AI Evaluation

3.9
Expert Rating

A genuinely free, MIT-licensed open-weights family with a novel self-generated training curriculum, scoring competitively with Claude Opus 4.8 on some agentic and coding benchmarks, though every figure is self-reported and there's no published safety evaluation yet.

Pros

  • Free, MIT-licensed across all three sizes with no revenue threshold
  • Novel self-improvement loop with explicit validity, difficulty and novelty gating
  • 9B model runs on a single consumer GPU or phone, an accessible entry point

Cons

  • All benchmark figures are DeepReinforce's own, not yet independently reproduced
  • No published system card, CBRN evaluation or safety documentation
  • 397B flagship needs multi-GPU, datacentre-class hardware to run at useful speed