Ornith-1.5-122B claims 76.9 on SWE-bench Pro vs 43.9 for Qwen3.8-Flash-Next
BerryBlush_Rogers691 · reddit · 2026-10-22
Open model Ornith-1.5-122B draws attention: its model card reports 76.9 on SWE-bench Pro vs 43.9 for Qwen3.8-Flash-Next, while slightly behind on Terminal-Bench 2.1 (81.2 vs 83.4). All results are the team's own.
- MoE reasoning model with tool calling and 256K context
- Available on Ollama; Q4KM is 77GB before runtime memory overhead
- Trained via generated tasks, agent setups and solutions used for RL
The poster asks how it handles long tasks on unfamiliar repos — recovering from failed tool calls, checking its own changes, and finishing the job — and whether it would displace current local coding models.
More from coding & agent
- PHP hit by two unauthenticated RCE flaws; ProjectDiscovery launches free internet-wide alert service — cyb3rops · 2026-10-23
- Modal open-sources Spindle, a Tinker-compatible training server that charges GPU time instead of tokens — josh_wills · 2026-10-23
- Testing the Azure Cosmos DB Agent Kit: what it found reviewing a Claude-built app, in RUs saved — adnan_hashmi · 2026-10-23
- Dev ditches endless chat threads for a visual dashboard to manage her AI agents — Salmaaboukarr · 2026-10-23
- Warp cuts coding agent verification bottleneck with per-agent computer-use and clip reviews — vikvang1 · 2026-10-23
- Computer Use Is Solved; Specifying Goals Is the New Bottleneck for Agents — alishbaimran_ · 2026-10-23