Ornith-1.5 Release: 397B MoE Self-Improvement Targets Claude Opus

The open-source Ornith-1.5 model family was officially released, featuring 9B Dense, 35B MoE, and 397B MoE configurations to cover various inference needs. Official data shows the flagship 397B model scored 86.1 on Terminal-Bench 2.1 and claims to rival Claude Opus 4.8 in reasoning, agent, and coding tasks. Licensed under MIT, the series features end-to-end self-improvement; the multimodal 35B-A3B topped Hugging Face trends, and all models are available on Ollama.

Confirmed

• The family includes 9B Dense, 35B MoE, and 397B MoE, released by DeepReinforce (Ornith Lab) under the MIT license.
• The 397B model scored 86.1 on Terminal-Bench 2.1, with claims of rivaling Claude Opus 4.8 in reasoning, agent, and coding tasks.
• The 35B-A3B is a multimodal model based on Qwen2.5-MoE, supporting image-to-text generation, and has topped Hugging Face trends.
• The core technology is an end-to-end self-improvement strategy where models autonomously propose tasks, generate scaffolds, and iterate via reinforcement learning.
• All models are available on Ollama, and AtomicChat has released quantized versions for the 9B and 35B models.

Unconfirmed

• Claims of "rivaling Claude Opus 4.8" and "targeting GPT-4.5" rely mainly on self-evaluation, lacking independent third-party verification.
• Demos of the 9B model running full agents offline lack first-hand review support.
• Comparisons stating the 35B MoE surpasses Qwen 3.6-35B and Gemma 4-31B lack original benchmark data support.

Why it matters

• If validated, the end-to-end self-improvement paradigm represents a new direction in AI training, potentially reducing reliance on manually constructed data.
• The combination of the MIT license, a full range from 9B to 397B, and rapid support from Ollama and the community provides a powerful open-source alternative to top-tier closed-source models.

2026-08-19 ~ 2026-08-20 · 24 related posts

• [source] Ornith 1.5 released: 9B dense and 35B/397B MoE models — tarruda · 2026-08-19
• Open-Source Ornith-1.5 Matches Claude Opus in Reasoning and Coding — aftahi_ai · 2026-08-19
• Open-source Ornith-1.5 released under MIT, claims Claude Opus-level performance — iamfakhrealam · 2026-08-19
• Open-source Ornith-1.5 hits SOTA with self-improving strategies — aftahi_ai · 2026-08-19
• Ornith-1.5: Self-Improving Models Match Claude Opus Performance — rohanpaul_ai · 2026-08-19
• Ornith-1.5 open-source family launches, 397B MoE claims Claude Opus 4.8 parity — KokaOP · 2026-08-19
• Ornith-1.5-35B Outperforms Qwen and Gemma in Reasoning and Coding — alejandroll10 · 2026-08-20
• Ornith 1.5 Released: MIT-Licensed 9B/35B Models with Quantization Benchmarks — Fun-Meaning-6474 · 2026-08-20
• Ornith-1.5 claims Claude Opus 4.8 performance — alexcovo_eth · 2026-08-20
• [source] Ornith-1.5-35B-A3B Trends on HF: Qwen2.5-MoE Multimodal Chatbot — ornith-ai · 2026-08-20

14 near-duplicate retellings: aftahi_ai · aftahi_ai · testingcatalog · Shruti_0810 · yacineMTB · Teknium · aftahi_ai · alexcovo_eth · aftahi_ai · rohanpaul_ai · ollama · nikola_mr64990 · nikola_mr64990 · Aiden_Tech_Ai