Ornith-1.5-122B: 122B open-weight MoE for coding agents, MIT licensed

Ornith's newest Ornith-1.5 family member: a 122B MoE built on Qwen3.5-122B-A10B that runs on a single 80GB GPU, with self-reported 84.5% SWE-bench Verified and 81.2 Terminal-Bench 2.1.

Why it matters

What you could build with it

Does it hold up?

Too early to judge: all benchmark figures are vendor-reported with no independent evaluation of the 122B variant yet, though the rapid appearance of community quants (AtomicChat, Unsloth) and detailed serving recipes suggest real deployability. Prior community runs of the smaller 35B model showed mixed results outside the vendor harness.

Built with Ornith-1.5-122B

• AtomicChat publishes imatrix-calibrated GGUF quants with full methodologyhf · Third-party quantization with importance-matrix calibration on the BF16 weights, per-tensor quantization rules, KL-divergence eval harness, and reproducible build commands — the kind of community packaging that makes a new model actually runnable.
• Ornith-1.5: does it really beat Claude Opus 4.8?article · Independent analysis of the Ornith-1.5 family's self-improvement training claims, checking the vendor benchmarks against community runs and the HN discussion — useful skepticism for anyone evaluating the 122B variant's numbers.
• unsloth/Ornith-1.5-122B-GGUFgithub · Unsloth GGUF builds referenced in the official model card for fast local inference and fine-tuning via the Unsloth library.
• AtomicChat/Ornith-1.5-122B-GGUFgithub · Community GGUF quants (AD-Q4_K-IQ4_XS and others) with documented importance-matrix calibration and KL-divergence evaluation.

Learn more

official docs →

First spotted on hf: source.

More AI models