Ornith-1.5 adds 122B: Based on the Qwen 3.5 architecture, some coding self-tests lead Qwen3.8-Flash-Next

Beating AI News Flash: The AI team DeepReinforce, founded by renowned NLP scholar Li Jiwei, has added a 122B MoE model to the Ornith-1.5 series.

The series was first released in August this year, initially including three tiers: 397B, 35B-A3B, and 9B. This time, the mid-range specification has been added. The new model has already opened its weights and is available in versions such as FP8, NVFP4, GGUF, and MLX.

Ornith-1.5-122B follows the model architecture of Qwen3.5-122B-A10B, with approximately 122B total parameters and about 10B activated parameters per inference. It supports a 256K native context, image understanding, tool calling, and reasoning.

Official self-testing shows that the 122B scored 84.5 on SWE-bench Verified (non-hack), higher than Qwen3.8-Flash-Next's 82.8. CyberGym was 80 versus 78, NL2Repo was 51.2 versus 45.2, and SWE-bench Pro (no-hack) reached 76.9 versus 43.9. However, it still trails on Terminal-Bench 2.1 at 81.2 versus 83.4.

It is worth noting that Qwen officially previously used Claude Code Harness and scored 62.5 on SWE-bench Pro, higher than Ornith's 43.9 measured using OpenHands. The conditions of the two tests are not entirely consistent.