Ornith-1.5-122B: Specs, License & How to Run

Ornith-1.5-122B is a Mixture-of-Experts coding and agent model from Ornith. Official hosts we opened on 2026-10-22: the ungated card ornith-ai/Ornith-1.5-122B (created 2026-10-16T07:14:27Z; 129 likes / 742 downloads), siblings Ornith-1.5-122B-FP8 (11 likes / 142 downloads) and Ornith-1.5-122B-GGUF (37 likes / 7,699 downloads), and the blog Ornith-1.5: From Self-Scaffolding to Self-Improvement. github.com/ornith-ai/Ornith-1.5-122B 404’d.

This page hangs off ChatGPT alternatives, closed LLM APIs, and Claude Code alternatives. That mapping is editorial. We have no sourced quality comparison against ChatGPT, Claude Code, or DeepSeek-V4.1-Flash. Head-to-head facts: Ornith-1.5-122B vs DeepSeek-V4.1-Flash, Flint vs Ornith-1.5-122B, Sage-A1.5 vs Ornith-1.5-122B.

The card calls it “the most capable and efficient member of the Ornith-1.5 family — a 122B MOE model designed for efficient single-GPU deployment,” with a quantized Ornith-1.5-122B-Mobile variant named for edge devices. Those adjectives stay vendor. Hugging Face pipeline_tag is text-generation; tags also include qwen3_5_moe.

What we verified (2026-10-22)

Field	Value	Source
Host (BF16)	huggingface.co/ornith-ai/Ornith-1.5-122B (not gated)	HF API
Created / last modified	2026-10-16T07:14:27Z / 2026-10-21T08:29:30Z	same
Likes / downloads	129 / 742	same
License tag	mit	HF cardData
LICENSE file fetch	resolve/main/LICENSE 404	this check
Pipeline tag	text-generation	HF API
Architecture tag	qwen3_5_moe	HF tags
safetensors.parameters	125,086,497,008 (~125.1B)	HF API
usedStorage (BF16)	250,188,099,758 bytes (~233 GiB)	same
FP8 sibling	created 2026-10-20; 11 likes; usedStorage 132,964,475,222	HF API
GGUF sibling	created 2026-10-20; 37 likes / 7,699 downloads	HF API
Native context	262,144 tokens	card
Serving named	Transformers ≥ 5.8.1; vLLM ≥ 0.19.1; SGLang ≥ 0.5.9	card
GitHub weights repo	404	GitHub API

The card name is 122B. Hugging Face’s parameter sum is ~125.1B. We keep both quotes and do not invent a reconciliation.

How to run (from the card, not tested)

The card says Ornith-1.5-122B is a reasoning model: the assistant turn opens with a think block, and the recipes enable a reasoning parser (reasoning_content) plus a tool-call parser for <tool_call> → OpenAI-style tool_calls.

Recommended sampling (copied):

Task	temperature	top_p	top_k	presence_penalty
General	1.0	0.95	20	1.5
Precise coding	0.6	0.95	20	0.0

Both rows also list min_p=0.0 and repetition_penalty=1.0.

vLLM (Apache-2.0) recipe from the card:

vllm serve ornith-ai/Ornith-1.5-122B \
  --served-model-name Ornith-1.5-122B \
  --host 0.0.0.0 --port 8000 \
  --max-model-len 262144 \
  --gpu-memory-utilization 0.90 \
  --enable-prefix-caching \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_xml \
  --reasoning-parser qwen3 \
  --trust-remote-code

--trust-remote-code executes repository code. Read the repo first or pin a revision. forkalt has not run this.

SGLang recipe from the card (--tool-call-parser qwen3_coder, --reasoning-parser qwen3, --context-length 262144). This site does not yet have a dedicated SGLang tool page. DeepInfer is DeepSeek’s engine and, in the 2026-10-21 README check, names DeepSeek-V4.1-Flash only — not Ornith. Compare DeepInfer vs vLLM.

Other routes named on the card: ollama run ornith-1.5:122b; llama-server against ornith-ai/Ornith-1.5-122B-GGUF; OpenAI-compatible clients (Hermes, OpenClaw, OpenCode) pointed at http://localhost:8000/v1. Those are copied labels.

For windows beyond 262,144 tokens the card suggests YaRN (factor 4.0 ≈ 1M tokens) and warns that open-source runtimes apply the scale statically, which can hurt ordinary-length quality. Enable it only when the workload needs it.

License

Card tag MIT. Same tag on the FP8 and GGUF siblings. The published license_link is the card’s LICENSE blob; a raw fetch of that path 404’d on 2026-10-22. Not legal advice.

What we did not invent

• VRAM, tokens/s, and “efficient / SOTA” stay vendor.
• Card table rows (SWE-bench, Terminal-Bench 2.1, CyberGym, HLE, WideSearch, …) stay self-reported.
• We do not rank Ornith against ChatGPT, Claude Code, or DeepSeek.

Pixel editors are a different job: PhotoCraft vs GIMP and the docs-based PhotoCraft install outline. Related models: Flint, Sage-A1.5, DeepSeek-V4.1-Flash, Reflection Beam, Ling-3.1-flash, Kimi K3. Serving: vLLM. UIs: Open WebUI, LibreChat, Jan.