Ornith 1.5 122B on BenchGecko · open weights under MIT for a single 80GB GPU
Ornith published open weights for Ornith 1.5 122B on 2026-10-20 under MIT. BenchGecko lists it with a 262,144 token context and records Ornith's 84.5 on SWE-bench Verified as a vendor claim.
What launched
Ornith released the weights of Ornith 1.5 122B on Hugging Face, with FP8, NVFP4, GGUF and MLX builds published on 2026-10-20. It is a mixture of experts reasoning model that Ornith says serves on a single 80GB GPU. It joins the Ornith 1.5 family that launched in August 2026 at 397B, 35B and 9B.
Ornith 1.0 was built on Qwen3.5 and Gemma 4 with continued pretraining, mid training and post training. Ornith 1.5 extends it with a self improvement loop that generates its own tasks, scaffolds and solution rollouts for reinforcement learning. BenchGecko added the model and the new Ornith provider today.
Specs
From the Hugging Face model card and checkpoint files:
- Parameters · about 125B in the BF16 checkpoint, named 122B
- Architecture · Qwen3.5 MoE layout, 48 layers, 256 experts with 8 active per token
- Context · 262,144 tokens, stored as 262,144 on BenchGecko
- Longer context · YaRN with factor 4 stretches the window to roughly 1M tokens
- Reasoning · thinking block on by default
- Weights · open, MIT licence
- Builds · BF16, FP8, NVFP4 at 77.7 GB, GGUF and MLX
- Runtimes · vLLM 0.19.1 or later, SGLang 0.5.9 or later, Transformers 5.8.1 or later, Ollama
Pricing on BenchGecko
Ornith publishes no first-party API price for this model, and no OpenRouter endpoint is live. BenchGecko leaves the price fields empty until a provider lists it.
For reference, Qwen3.5 122B A10B, released 2026-02-25, lists at $0.26 input and $2.08 output per M tokens with the same 262,144 token context.
Vendor-reported benchmarks
Ornith reports these results on the model card, averaged over five runs with anti reward hacking safeguards. BenchGecko records them as vendor claims. No independent scores or Gecko Tests are attached yet.
- SWE-bench Verified · 84.5
- SWE-bench Pro · 76.9
- SWE-bench Multilingual · 76.4
- Terminal-Bench 2.1 with Claude Code · 81.2
- CyberGym · 80.0
- NL2Repo · 51.2
- HLE with tools · 48.0
- HLE without tools · 37.68
- SciCode · 53.8
- BrowseComp · 78.5
- WideSearch · 75.2
- MCP-Atlas · 74.6
- Toolathlon-Verified · 70.8
- ClawEval · 79.4
On the same card, Ornith lists Nemotron 3.5 Super VL at 61.2 on SWE-bench Verified and 66.7 on Terminal-Bench 2.1.
Sources
Source: BenchGecko
Written automatically from BenchGecko's tracked data for this period; every number comes from the changelog, model, mindshare and pricing datasets. Checked against the source data before publishing.