Ornith-1.5 122B Tested: Vision, Coding, and Reasoning Up Close
Hands-on tests of Ornith-1.5 122B cover image recognition, single-file HTML animation, social reasoning, and translation across 80 languages.
What is Ornith-1.5 122B?
Ornith-1.5 is an open weights mixture of experts model with about 122 billion total parameters, released under the MIT license, meaning anyone can download it, run it locally, and build commercial products on top of it without restriction. It reasons before answering, accepts both images and text as input, and handles a context window of roughly 260,000 tokens. The model is pitched by its creators as “stronger, faster, and leaner” than its predecessor, aimed squarely at coding, agentic workflows, and tool use rather than general chit chat.
TL;DR
- Ornith-1.5 is a 122B parameter mixture of experts model released under the MIT license, built for coding, agents, and tool use rather than casual conversation.
- The model supports vision input, correctly identifying a dish as Mughlai curry from a photo and backing it up with a ten-step recipe and ingredient callouts drawn directly from the image.
- In a debugging test on a mock legal application, the model found and fixed a single faulty line governing inheritance rules without being told where to look.
- A single HTML file animation test showed the model generating a full kangaroo tattoo illustration with shading and outline sequencing, with no external libraries or assets.
- On a social reasoning puzzle involving a WhatsApp screenshot, the model correctly parsed ambiguous context (a work deploy mistaken for an affair) and drafted a tactful, funny resolution.
- Translation across 80 languages was largely accurate for major world languages, but quality dropped noticeably for low resource languages like Thai, Hawaiian, and Zulu.
- Benchmark claims put Ornith-1.5 ahead of similarly sized models on agentic and coding tests, and competitive with models several times its size on SweepBench Pro.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
How was the vision test set up?
The test gave Ornith-1.5 a photo of a curry dish with no text hints, asking it to name the dish, identify its regional origin, and write a complete recipe. This checks whether the model actually parses visual detail rather than producing a generic guess.
The model’s reasoning process, visible before the final answer, weighed several possibilities: Thai red curry, Caribbean curry, and a general Indian curry, before settling on an answer. It noticed a specific visual detail, an oddly sliced ingredient resembling sausage, and used that to refine its judgment. It landed on Mughlai, a North Indian chicken and potato curry, correctly pointing out chunks of chicken, red chili oil, and other visible ingredients in the photo. It then produced a full recipe for four people, with quantities, ten clear steps, and additional cooking tips.
The reasoning was thorough, arguably more thorough than necessary for a single photo, which suggests the “leaner” claim doesn’t fully apply when extended thinking is enabled by default. That said, response speed was still fast, and the depth of reasoning translated into a more accurate final answer than a quicker guess likely would have produced.
Can Ornith-1.5 actually debug real code?
To test practical coding ability, the model was pointed at a full stack application called Legacy Vault, a tool modeled on software law firms use to calculate inheritance after someone dies. The application contained a logic bug: a legal survivorship rule requires a beneficiary to outlive the deceased by 30 days to inherit, but the app had the logic reversed. In the test scenario, a beneficiary who had died 132 days before the person whose estate was being settled was still shown receiving a payout.
The model was given the goal with no hints about where the bug lived in the codebase. Using an agent setup, it worked through the application, flagged some false leads along the way, and eventually isolated the single line causing the inversion. After the fix, the application correctly zeroed out the deceased beneficiary’s share and correctly allocated funds to another beneficiary who had outlived the required period. The model found the issue without hand holding and the application’s ledger balanced afterward, which is a meaningful marker of practical coding capability beyond passing a static benchmark.
How does it handle creative coding tasks?
A creative coding prompt asked the model to produce a single HTML file animating a kangaroo tattoo being drawn on a thigh, with no external dependencies. This tests visual design judgment, animation sequencing, and whether the model can deliver a complex, cohesive result in one attempt.
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
The output built the tattoo in stages: an outline first, then shading, then a filled silhouette, finishing with additional linework and background detail like scattered stars. The style resembled traditional tattoo art with solid black fills and a limited color palette. It’s not a substitute for image generation tools, but as a self-contained HTML and CSS/JS animation with no assets pulled from outside the file, the output held together well and ran cleanly at different playback speeds.
Is Ornith-1.5 good at reasoning beyond code?
Two tests pushed the model on reasoning rather than generation. The first was a physics style image prompt showing people holding different wires, asking which ones would get shocked. The model correctly identified the two people who would complete a circuit through their bodies, explained that a shock requires a closed path, and summarized its reasoning in a table, without any accompanying text describing the scenario beyond the image itself.
The second test was a social reasoning puzzle: a WhatsApp screenshot involving a boss messaging about an 11pm deploy, phrased ambiguously enough that a suspicious spouse reading only part of the conversation might misread it as an affair. The model correctly parsed who said what, recognized “tonight” referred to a software deployment rather than anything personal, and understood that the spouse had only seen a partial exchange. It then produced a sequenced response: reassure the spouse first with a clear, non-defensive message, respond to the boss about the deploy, and propose a believable, low-drama plan for the evening. The reasoning and the writing in the response were coherent and genuinely well-constructed for a model being tested on logic rather than prose.
How well does it translate across languages?
The translation test asked the model to render a single sentence (a lawn-mowing neighbor pleading for mercy while distracted by someone on a trampoline) into 80 different languages in their native scripts. The model completed all of them in a single pass, one line per language, though it didn’t preserve the sentence exactly as given in every case.
Major world languages, including Spanish, French, German, Korean, Japanese, Hebrew, and Arabic, read correctly. Lower resource languages fared worse. Thai, Hawaiian, and Zulu translations looked unreliable, and regional South Asian languages along with some Southeast Asian languages also showed quality drops. This pattern is common across large multilingual models: strong coverage for widely spoken languages, weaker and sometimes fabricated-looking output for languages with less training data available.
Is Ornith-1.5 worth running locally?
For teams already set up to run open weight models with tools like llama.cpp, Ornith-1.5 is a reasonable candidate for coding and agentic tasks specifically. It located and fixed a non-trivial logic bug in a multi-component application without guidance, handled multiple vision reasoning tasks correctly, and produced competent creative code output in a single file. Benchmark comparisons cited by its creators place it ahead of similarly sized models on coding and agentic tests, and competitive with models several times larger on real-world software issue benchmarks like SweepBench Pro.
It is not uniformly polished. Its reasoning traces can run long even on simple inputs, its translation quality is uneven outside major languages, and its humor and nuance in casual tasks is solid without being exceptional. For a model being positioned as a daily working tool for coding and agent use rather than a flagship chat assistant, that trade-off looks reasonable.
Frequently Asked Questions
What is Ornith-1.5 122B designed for?
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
It’s built primarily for coding, agentic workflows, and tool use, rather than general purpose chat. It supports both text and image input and reasons through problems before producing an answer.
Is Ornith-1.5 open source?
Yes. It’s released under the MIT license as an open weights model, meaning it can be downloaded, run locally, and modified or built upon without licensing restrictions.
How large is Ornith-1.5 and what hardware does it need?
It’s a mixture of experts model with about 122 billion total parameters. VRAM requirements depend heavily on the quantization format chosen and the context length used; running it with a long context window increases memory use substantially.
How good is Ornith-1.5 at translation?
It handles major world languages like Spanish, French, German, Japanese, and Arabic well, but accuracy drops for low resource languages such as Thai, Hawaiian, and Zulu, and for some regional South Asian and Southeast Asian languages.
Can Ornith-1.5 understand images, not just text?
Yes. In testing, it correctly identified a specific dish from a photo with accurate ingredient detail, solved a visual physics reasoning puzzle about electric shock risk, and read a chat screenshot accurately to resolve an ambiguous social situation.
