Ornith has released Ornith-1.5-122B, a 122B mixture-of-experts reasoning model, with open weights under the MIT license. The company announced the model on X at 22:36 Beijing time on October 20 and published the weights on Hugging Face.
According to the model card, the model serves on a single 80GB GPU, and the serving commands can be extended to shard it across more GPUs. The announcement on X describes the hardware as a single NVIDIA DGX card. The model card does not state how many parameters are active per token.
The model card says the model handles context windows of up to 262,144 tokens. Ornith suggests YaRN scaling for longer tasks and says a scaling factor of 4.0 extends the usable window to roughly 1M tokens, while noting that static scaling can slightly reduce quality on ordinary-length inputs.
As a reasoning model, Ornith-1.5-122B opens each answer with a reasoning block by default. The model card lists Transformers 5.8.1, vLLM 0.19.1 and SGLang 0.5.9 or later as the required runtimes, and its examples also cover GGUF builds through llama.cpp and Unsloth.
Ollama said at 23:22 Beijing time on October 20 that the model is available on Ollama and can be started locally with the command ollama run ornith-1.5:122b.
Ornith's own benchmark table lists 84.5 on SWE-bench Verified, 76.9 on SWE-bench Pro and 78.5 on BrowseComp. These are Ornith's results and have not been independently verified. The model card says the SWE-bench runs used the OpenHands harness with safeguards against reward hacking.