Ornith-1.5: Self-Improving Models Match Claude Opus Performance
rohanpaul_ai · x · 2026-08-19
Ornith-1.5 introduces a major step toward foundation models built through end-to-end self-improvement, extending the self-scaffolding framework into a complete loop where the model proposes tasks, generates scaffolds, and produces reinforcement learning solutions.
Model Specs & Performance:
- 397B MoE: Scores 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, matching Claude Opus 4.8 and outperforming GLM-5.2 and DeepSeek-V4-Flash.
- 35B MoE: Activating only 3B parameters per token, it significantly outperforms Gemma 4-31B and Meta Muse Glimmer-30B on agentic coding tasks.
- 9B Dense: A mobile-quantized version runs on phones, substantially outperforming larger models like Gemma 4-31B and Qwen 3.6-35B.
More from Models
- Ling-3.0 Open Source: 6 Base Checkpoints Released with WSM Training Method — bclavie · 2026-08-19
- Grok 4.6 lands on Amazon Bedrock with 500k context and configurable reasoning — SpaceXAI · 2026-08-19
- empero-ai's Qwen3.8-9B-Distill trends on Hugging Face — empero-ai · 2026-08-19
- z-lab's Qwen3.8-27B DFlash2: block-diffusion draft model for fast inference — z-lab · 2026-08-19
- GGUF quantized Qwen3.8-9B-Distill lands for llama.cpp local runs — empero-ai · 2026-08-19
- Replit Free Mode powered by GPT-5.6 Luna: $20/mo enough to deploy 10+ apps — amasad · 2026-08-19