← sources

Ornith-1.5-122B: anyone tried it in a coding agent yet?

source: reddit (reddit:LocalLLaMA)published 2026-10-21 16:53ingested 2026-10-22 02:42

retained text

The [model card](https://huggingface.co/ornith-ai/Ornith-1.5-122B) has a pretty noticeable result: **76.9 on SWE-bench Pro**, compared with 43.9 for Qwen3.8-Flash-Next in their evaluation. On Terminal-Bench 2.1 it's slightly behind, at 81.2 vs 83.4. These are the team's own results, but the difference between those two comparisons is interesting.

It's a MoE reasoning model with tool calling and a 256K context window. There's an [Ollama release](https://ollama.com/library/ornith-1.5:122b) already; the Q4\_K\_M package is around 77GB, before runtime memory overhead.

Their [training approach](https://ornith.ai/ornith_1_5.html) involves generating tasks, the agent setups used to solve them, and solutions, then using those for reinforcement learning. I'm curious how much of that shows up when you give it an unfamiliar repo and your own tools.

Has anyone put it through a longer coding task? Mostly interested in whether it needs less babysitting: recovering from a failed tool call, checking its changes, and actually finishing the job. What would make you switch from whichever local model you're using now?

document fetched 2026-10-22 00:44 · canonical https://reddit.com/r/LocalLLaMA/comments/1xbslby/ornith15122b_anyone_tried_it_in_a_coding_agent_yet