Ornith-1.5-122B fits an open MoE powerhouse on a single 80GB GPU

Ornith-1.5-122B is an open-source mixture-of-experts (MoE) reasoning model from ornith-ai built specifically for hard coding, multi-step agent workflows, and deep reasoning.

While Ornith-1.0 leaned on continuing pretraining atop Qwen3.5 and Gemma4, this new release shifts to a full self-improvement loop. Instead of training on static human-curated datasets, it continuously generates synthetic problem sets, builds execution scaffolds, and optimizes its rollout strategy using reinforcement learning.

Specs and benchmark gains

The model features 122B total parameters with roughly 10B active parameters per token. The standout hardware win is efficiency: despite its scale, the full MoE checkpoint serves on a single 80GB GPU. For edge use cases, the team also prepared a quantized mobile variant called Ornith-1.5-122B-Mobile.

Across major benchmarks, it frequently punches above Qwen3.8-Flash-Next (125B). On SWE-bench Pro under anti-hacking settings, Ornith-1.5-122B scores 76.9 versus Qwen's 43.9. On SWE-bench Verified, it hits 84.5 compared to 82.8.

In tool-assisted reasoning on Humanity's Last Exam (HLE), it posts 48.0 against Qwen's 38.0, and reaches 51.2 on NL2Repo. However, it slightly trails Qwen3.8-Flash-Next on TerminalBench 2.1, scoring 81.2 against 83.4.

Because it functions as a pure reasoning model, it returns its chain-of-thought within a think block by default while offering native OpenAI-compatible tool call parsing.

Getting it running

The weights are available now on Hugging Face, alongside sample notebooks for Google Colab and Kaggle and Docker runner support.

Serving locally requires recent inference runtimes: Transformers 5.8.1+, vLLM 0.19.1+, or SGLang 0.5.9+. You can spin up an OpenAI-compatible endpoint with a standard vLLM serve command.

Context length tops out at 262,144 tokens out of the box. If you need extreme context, you can apply static YaRN RoPE scaling with a 4.0 factor inside vLLM or SGLang to push the context to roughly 1 million tokens. Keep in mind that static YaRN slightly impacts output quality on short prompts, so leave it off unless required.

Why it matters

For developers and teams building self-hosted coding agents, Ornith-1.5-122B hits an undeniable sweet spot. Squeezing a top-tier 122B MoE into a single 80GB GPU eliminates the latency and operational complexity of tensor-parallel multi-GPU setups.

It also highlights how powerful closed-loop reinforcement learning has become: by generating and scoring its own tasks, a mid-sized open model can match or beat proprietary lab offerings in agentic benchmarks. The main caveat remains inference latency during long chain-of-thought traces, but for tough SWE tasks, it is an easy recommendation to test.