DeepReinforce has Ornith-1.5 added a new 122B MoE model and released its weights
BIBIBI
AT A GLANCE
DeepReinforce has Ornith-1.5 added a new 122B MoE model, claiming that it leads in some programming self-tests Qwen3.8-Flash-Next。
Article
Ornith-1.5 New addition 122B: based on the Qwen 3.5 architecture, leads in some programming self-tests Qwen3.8-Flash-Next Founded by renowned NLP scholar Li Jiwei, the AI team DeepReinforce has added a new Ornith-1.5 model to the 122B MoE series.
The series was released as early as August this year, when it included 397B、35B-A3B and 9B tiers; this release fills in the intermediate specification. The new model has released its weights and is available in FP8、NVFP4、GGUF and MLX versions.Ornith-1.5-122B retains the model architecture of Qwen3.5-122B-A10B, with approximately 122B total parameters and approximately 10B parameters activated per inference. It supports 256K native context, image understanding, tool calling, and reasoning.
Official self-tests show that 122B scored SWE-bench Verified on 84.5(non-hack), higher than Qwen3.8-Flash-Next's 82.8. On CyberGym, 80 versus 78，NL2Repo was 51.2 versus 45.2，SWE-bench Pro (no-hack) reached 76.9 versus 43.9. However, on Terminal-Bench 2.1 still trailed 81.2 by 83.4.
Notably, Qwen officials previously used Claude Code Harness and achieved SWE on 62.5-bench Pro, higher than the 43.9 scored by Ornith using OpenHands; the conditions of the tests were not fully consistent.
Original link https://m.theblockbeats.info/flash/372992
Key points
01
DeepReinforce has Ornith-1.5 added a new 122B MoE model.
02
The report says the series previously included 397B、35B-A3B and 9B three tiers.
03
The new model has released its weights and is available in FP8、NVFP4、GGUF and MLX versions.
04
Ornith-1.5-122B It retains Qwen3.5-122B-A10B's model architecture, with approximately 122B total parameters and approximately 10B parameters activated per inference.
05
The model supports 256K native context, image understanding, tool calling, and reasoning.
06
In official self-tests,Ornith-1.5-122B scored SWE-bench Verified on 84.5, while Qwen3.8-Flash-Next scored 82.8.
07
In official self-tests, CyberGym was 80 versus 78，NL2Repo was 51.2 versus 45.2，SWE-bench Pro (no-hack) was 76.9 versus 43.9。
08
Terminal-Bench 2.1 In the test,Ornith-1.5-122B trailed 81.2 by 83.4.
09
The report notes that Qwen officials previously used Claude Code Harness on SWE-bench Pro and achieved 62.5, while Ornith, using OpenHands, scored 43.9; the test conditions were not fully consistent.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
This is Ornith-1.5 a new intermediate-specification model in the series.MoE The model does not use all parameters in every inference: it has approximately 122B total parameters, but activates approximately 10B each time. The performance advantages in the report come from official self-tests, and different tests may have used different tools and conditions, so the scores should not be treated directly as final conclusions under a unified environment.
Why it matters to readers
The new model has released multiple weight versions and supports long context, image understanding, and tool calling; if retests under consistent conditions continue to show advantages on some programming benchmarks, it could provide a new reference for selecting relevant models.
Beginners should first consider whether the model is easy to deploy, the required hardware, and performance on real-world tasks, rather than judging it solely by a single leaderboard score.
Risks and unknowns
The performance data primarily come from official self-tests; the evidence does not provide independent retest results.
The testing tools and conditions in some comparisons are not fully consistent, so the scores may not be directly comparable.
The evidence does not specify the hardware requirements or costs for deploying this model.
The evidence does not provide an official model page, weight repository, or complete technical documentation; the related release information still requires verification.
Related Developments
Loading event timeline…
Related concepts
MoE
This term is not in the glossary yet. Browse related concepts in the glossary.
