Ornith-1.5-122B: MIT-Licensed Open Model That Fits One 80GB GPU

Ornith has put out Ornith-1.5-122B, an MIT-licensed mixture-of-experts model that the Hugging Face model card says runs on a single 80GB GPU. The headline number is 84.5 on SWE-bench Verified, with 80.0 on CyberGym and 81.2 on TerminalBench 2.1. Those scores are the vendor's own, and the card itself notes custom harnesses and a modified chat template. This post gives you the full table, the commands to run it, the quantization options, and a clear line between what is claimed and what is checkable.

The model repository appeared on Hugging Face on October 16, 2026, with FP8, NVFP4, GGUF and MLX builds following on October 19 and 20, based on the Hub listings. For the family's history, see our posts on Ornith-1.0's self-scaffolding approach and Ornith-1.5 and whether it is really self-improving.

TL;DR: Ornith-1.5-122B at a glance

The benchmark table: vendor numbers, side by side

This is the model card's own comparison. Every number is reported by Ornith; none are independently reproduced as far as we found.

The card publishes more rows (SciCode, WideSearch, ClawEval, HLE without tools); we kept the most decision-relevant ones. The pattern: Ornith leads the coding and security rows against similarly sized open models, while DeepSeek-V4-Flash and Qwen3.8-Flash-Next are ahead on several terminal, browsing and HLE rows. It is strong for its size, not a clean sweep.

The caveats the card admits

• The Terminal-Bench run required modifying the Qwen chat template.
• The SWE-bench results use anti-hacking safeguards ("non-hack"), which is good practice but is also a custom setup.
• Several evaluations use custom harnesses or judge models.
• There is no limitations section, and the training method is described only by the citation title, "From Self-Scaffolding to Self-Improvement."

The gap between SWE-bench Pro (76.9) and the Qwen3.8-Flash-Next figure (43.9) is unusually large and worth reproducing before you believe it. Our recent coverage of how benchmark claims can break includes the DeepSeek V4.1 Flash cheating report and the vendor-vs-verified framing of Nemotron 3.5 Super VL.

How to run it

Required serving versions are Transformers 5.8.1 or newer, vLLM 0.19.1 or newer and SGLang 0.5.9 or newer. From the model card:

vllm serve ornith-ai/Ornith-1.5-122B \
  --served-model-name Ornith-1.5-122B \
  --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

Other routes listed: SGLang with --tool-call-parser qwen3_coder, ollama run ornith-1.5:122b, llama-server -hf hf.co/ornith-ai/Ornith-1.5-122B-GGUF, and Docker Model Runner. Recommended sampling is temperature 1.0, top_p 0.95 and top_k 20 for general work, and temperature 0.6 for precise coding.

Note --trust-remote-code in the command: it executes code shipped with the repository, so read the repo first or pin a revision.

Quantization and local options

Why the ladder matters

At BF16 the weights are around 250GB (the Hub lists 125B parameters), so "one 80GB GPU" implies a quantized build such as FP8 or NVFP4. The card does not spell out which precision the single-GPU claim assumes, so check memory with your chosen format and context length. A long 262K context adds KV-cache memory on top of the weights.

Hub listings show the official FP8, NVFP4 and GGUF repos, plus community builds: Unsloth GGUF, AtomicChat GGUF, an EXL3 4-bit build, an MLX build from Osaurus, and a speculative-decoding drafter. A report we reviewed says AtomicChat's ladder spans 132.6GB down to 35.0GB so it fits 64GB Macs. Smaller quants trade accuracy for memory, and no one has published a quality curve for this model yet. For comparable local-AI trade-offs see our guides on Qwen3.8-Flash-Next in FP4 on RTX PRO 6000 and DGX Spark and Qwen3.8-Flash-Next 125B on a consumer GPU.

Ornith vs the rivals you would actually pick

"Better fit" here follows the vendor table and is not a measured ranking.

How to read these benchmarks without fooling yourself

Coding benchmarks are the most gamed numbers in open-model releases, so it helps to know what each row measures.

SWE-bench Verified asks a model to fix real GitHub issues in real repositories, judged by the project's own tests. It is public, which means training data can leak the answers; the "non-hack" label on the card suggests Ornith applied safeguards against shortcuts such as reading test files or git history, but the card does not describe them. SWE-bench Pro is harder and less saturated, which makes the 76.9 versus 43.9 gap against Qwen3.8-Flash-Next the number to scrutinize first. A gap that large can come from a real capability difference, but it can equally come from different scaffolding, step limits or retry budgets between the runs.

TerminalBench 2.1 measures agents working in a shell. The card says the run needed a modified Qwen chat template, which means the number depends on prompt formatting that you must replicate to get the same result. CyberGym tests vulnerability reproduction, where Ornith's 80.0 stands above every other column. HLE is a hard knowledge test where the larger DeepSeek and Beam models lead; a 122B model not winning there is what you would expect.

The most useful habit is to ask, for every score, who ran it, with which harness, with how many attempts, and whether anyone else has reproduced it. For Ornith the answer to the last question today is no.

What the family history tells you

Ornith's first release, 1.0, was pitched around self-scaffolding: models that build their own agent harness for coding. The 1.5 family in August added 397B, 35B and 9B sizes under an MIT license, and the community asked on Hugging Face discussions for a size between 35B and 397B. The 122B fills that gap, which explains the quick uptake from quantizers: Unsloth, AtomicChat, Osaurus and others published builds within days of the weights.

That ecosystem speed is a real signal that people want a mid-size, permissively licensed coding model. It is not a quality signal. Likes on a new repository measure curiosity, and the download counts on the official repository were still in the dozens when we checked the Hub.

A 30-minute test plan

If you want to decide for yourself rather than trust a table:

1. Pick ten issues from your own repository that were fixed by humans, with tests, and hide the fix commits.
2. Serve the model with the vLLM command above, using the recommended coding sampling of temperature 0.6.
3. Run the same ten tasks through the agent harness you already use, once with Ornith and once with your current model, with the same step and time budget.
4. Record pass rate, wall-clock time and tokens, and note failures that came from tool-call formatting rather than reasoning; the tool parser flags matter.
5. Repeat at the quantization you plan to ship.

Ten tasks is not statistics, but it exposes the most common failure, which is a model that scores well on public benchmarks and stumbles on your build system. The gap between a vendor table and your own repository is the real verification.

What to verify before you adopt it

1. Run your own repo tasks. SWE-bench Verified is a public benchmark, and contamination and harness effects are real.
2. Test at your quant. Run the exact FP8, NVFP4 or GGUF build you will deploy.
3. Check the long context. Static YaRN scaling can slightly hurt short-input quality, per the card.
4. Review the repo. --trust-remote-code plus a new organization deserves a quick audit.
5. Watch for independent evals from third parties over the next week; none had appeared when we checked.

What is still unknown

• The active parameter count and the training recipe.
• Independent reproductions of the headline scores.
• Which precision the single-GPU claim assumes, and the memory at full context.
• Whether the GGUF builds are complete; the card refers to the GGUF as something to be published, while the Hub already lists repos.

Related reading

• Ornith-1.0: self-scaffolding open models for agentic coding
• Ornith-1.5: is this open model really self-improving?
• Reflection Beam 501B open weights
• Nemotron 3.5 Super VL 120B
• Qwen3.8-Flash-Next FP4 locally
• Qwen3.8-Flash-Next 125B on a consumer GPU

Benchmarks are the vendor's, as listed on the Hugging Face model card on October 20, 2026.