How Does Ornith-1.5-122B Perform on Real Tasks? Hands-On Test Results

How does Ornith-1.5-122B actually perform on real tasks?

In one creator's hands-on tests, Ornith-1.5-122B did well on most of them. It found and fixed a planted inheritance-logic bug in a multi-container Docker app with no hand-holding. It identified a dish from a photo and wrote a full recipe. It read a chat screenshot accurately, solved an electric-shock diagram, and built an animated tattoo in one HTML file. Its weak spot was translation: low-resource languages came out poor.

What is Ornith-1.5-122B?

According to its model card, Ornith-1.5-122B is the most capable and efficient member of the Ornith-1.5 family, a 122B mixture-of-experts model designed for efficient single-GPU deployment. The model card says it extends Ornith-1.0, which was built on Qwen3.5 and Gemma4 with extra continued pretraining, mid-training, and post-training. Version 1.5 widens the self-improvement loop to jointly optimize task generation, scaffold construction, and solution rollouts.

The pitch is "stronger, faster, and leaner": shorter outputs, better instruction following, and stronger task decomposition. The card says it runs on a single DGX and lists a quantized Ornith-1.5-122B-Mobile variant for edge deployment.

How do the published benchmarks compare?

The card compares Ornith against Qwen3.8-Flash-Next (125B), Qwen3.8 (27B dense), Nemotron 3.5 Super VL (120B), DeepSeek-V4-Flash-0731 (284B), and Beam (501B). These are the vendor's numbers, not independent results. Some cells are blank for rivals.

Ornith does not win everything. DeepSeek and Beam, which are much larger, edge it out on a few agentic and reasoning rows. Its strongest showing is on real-issue software tasks, which matches what the bug-fix test checked.

Can it fix a real bug on its own?

This was the main test. The app, called Legacy Vault, is a Docker-based full-stack tool that law firms would use after a death. It reads a will, assets, debts, and family, then computes who inherits what. The will requires a beneficiary to outlive the deceased by 30 days. The app had that rule backwards.

The symptom was obvious. Thomas died 132 days before Margaret, yet the app paid him AUD 1.95 million. The creator ran the model inside an agent harness they had already installed. It got a goal with no step-by-step guidance.

The result:

1. The model found the single reversed line on its own.
2. It rebuilt the app.
3. After a hard refresh, Thomas received nothing, and Walter, who outlived Margaret, received his AUD 519,000 share.
4. The ledger still balanced.

It also flagged some false alarms rather than "fixing" them, and its verification step looked solid. It took a while but not an excessive one.

How good is its vision?

Two image tests stood out.

The curry photo. The prompt asked for the dish name, its origin, and a full recipe. The model's reasoning compared Thai red curry, Caribbean curry, and a general Indian curry. It noticed an odd sausage-like slice in the image and checked itself before committing. Its final answer was a Mughlai-style North Indian chicken and potato curry, with the visible chicken pieces and red chili oil cited as evidence. The recipe served four, with quantities, ten steps, and tips.

The reviewer's caveat is worth noting. The thinking went back and forth a lot on one photo, so it is not especially lean at its default reasoning level. The reviewer says you can reduce the thinking.

The wiring diagram. Given an image of people holding different wires, with no accompanying text, the model inferred the question and answered that A and D would be shocked. It explained that a shock needs a complete path through the body, and that D is shocked even on a stool because he touches both wires. It added a summary table.

Can it read a screenshot and reason about people?

The WhatsApp test was a deliberately awkward scenario: a boss, an 11:00 p.m. deploy, and a suspicious wife. The model read the screenshot correctly. It worked out who said what, that "tonight" meant the deploy and not something else, and that the wife had only seen the first two messages.

It then sequenced its advice sensibly: reply to the wife first, then the boss, then handle the deploy, with a reason for each order. One line in its draft to the wife, "It's a deploy, not a dossier," was the reviewer's favorite. The reviewer rated this the best answer of the session.

Can it handle creative coding in one shot?

The prompt asked for a single HTML file that animates a kangaroo tattoo being drawn on a thigh. The goal was to test visual design, animation code, and whether a complex result arrives in one go.

The file worked on the first open. The animation draws an outline, then shading, then angular details, in an Australian traditional style with solid black fills and a limited palette. Stars are scattered around the figure, and a speed control lets you run it faster. There are no external dependencies. The reviewer noted defects but called it strong for what is, in the end, plain HTML and not an image generator.

Is its multilingual translation any good?

This was the weakest test. The prompt asked for a short, humorous plea about a man mowing his lawn while a neighbor jumps on a trampoline, translated into 80 languages. Two problems showed up:

• The model changed the sentence instead of translating it exactly.
• It returned 81 languages, one line each, in native scripts.

The reviewer judged Spanish, French, German, Korean, Japanese, Hebrew, and Arabic as correct. Thai, Hawaiian, and Zulu looked invented. Regional Indian and Pakistani languages and several Southeast Asian ones also fell off. The reviewer is not a native speaker of all of these, so treat this as a spot check, not a measurement. The pattern is common across LLMs, but if you serve low-resource languages, test your own.

Is Ornith-1.5-122B worth trying?

For coding and agent work, the evidence says yes. It is open under MIT, fits the "daily worker" role the reviewer described, and its real-issue software results are its strongest suit, both on paper and in the Docker test. Expect long reasoning on vision tasks unless you tune it down. Expect uneven quality outside major languages. The context window is also on the smaller side at just over 260,000 tokens.

One test per category is not a benchmark. The bug in the Docker app was a single reversed condition, and harder multi-file defects may behave differently.