01What Ornith 1.5 122B is and how it works
Ornith 1.5 122B is a new open AI model with 122 billion parameters. It is a mixture of experts model, so it only uses a small part of itself on each task.
Julian Goldie explains this with a staffing agency. Instead of sending all 122 workers to do every job, the agency sends only the ones trained for that task. The model only activates a fraction of its parameters at a time.
That is how it fits on a single GPU while still being powerful.
The model comes with an MIT license and a self improving training loop. The release has a few key facts:
- 122 billion parameters, mixture of experts
- 256,000 token context window
- MIT license
- Single GPU deployment
In this episode Julian goes through the benchmark scores, the file sizes, how to run it with Ollama and what to do with it today.
02Ornith 1.5 122B benchmark scores
The scores for Ornith 1.5 122B are about agent work, coding and real tasks, not general chat. Here is what Ornith reports:
- SWE-bench Pro (the harder version of that coding test): Ornith scored 76.1%. Qwen 3.8 Flash, another top open model, scored 43.9% on the same test.
- Terminal-Bench 2.1, which tests hard tasks through command line tools: 81.2%.
- CyberGym, which covers security vulnerability tasks: 80%.
- BrowseComp, which tests how well it can research difficult questions across the web: 78.5%.
Julian makes a clear point about these numbers. They are reported by Ornith themselves, and third parties have not verified them yet.
Even so, early community feedback is strong. People are downloading and testing it now, and the model card is very detailed about how the evaluations were run. Julian calls it credible, but you should still test it on your own work.
03Context window and hardware needs for Ornith 1.5 122B
The context window is 256,000 tokens. Julian compares it to working memory. The more tokens a model has, the more it can hold in its head at once.
That could be a whole code base, a long research document or a detailed business process, all at the same time.
The hardware is the real limit. Here are the sizes:
- Full BF16 model: about 250 GB. This is full precision and will not run on a standard laptop.
- GGUF quantized version through Ollama: about 73 to 77 GB, depending on the version.
Ornith released quantized versions alongside the main model. Even at 73 to 77 GB, that is big, but it makes local testing possible on high end workstations.
The official deployment uses FP8 on a DGX. You will not match that performance with the Ollama version, but you get a real feel for the model. Julian is honest that this is not a consumer laptop model.
04How to run Ornith 1.5 122B with Ollama
The practical way in for most people is Ollama. The model is listed there as Ornith 1.5 with the 122B tag. Here is the path Julian suggests:
- Check that you have a high end GPU or workstation with enough memory for a 73 to 77 GB file.
- Open Ollama and find the Ornith 1.5 122B listing.
- Pull the model.
- Run it.
- Test it on a real task you actually need done, not a hello world prompt.
If you are on standard hardware, Julian suggests bookmarking the GGUF quantized versions. As quantization improves over the next few weeks, local performance should get better, and the community is already working on it.
Remember that the Ollama version will not give the same performance as the official FP8 deployment on a DGX. Treat your first run as a test of fit for your work, not a final verdict.
05What to do with Ornith 1.5 122B today
Julian gives clear next steps depending on where you are.
- You have a high end GPU: pull the model on Ollama and test it against a real workflow.
- You are on standard hardware: bookmark the GGUF quantized versions and check back as they improve.
- You are a business owner using AI automation: look at your current workflows. Which ones involve research, multi-step tasks or tool use?
Ornith 1.5 is built for research, multi-step tasks and tool use. With an MIT license, the barrier to putting it into a real workflow is low.
Good test tasks match the benchmarks. Try a coding fix, a command line job or a web research question. These are the same kinds of work the scores measure, so you can compare what you see with what Ornith reports.
Give it something real, check the output yourself, and keep notes on where it helps and where it does not.
06Tips and limits when testing Ornith 1.5 122B
A few cautions keep your testing honest.
- Treat the scores as claims. They are self-reported and not independently verified yet.
- Do not expect laptop use. The model is not for a consumer laptop, and even the quantized file is 73 to 77 GB.
- Expect a gap to the official setup. The Ollama version will not match the official FP8 deployment on a DGX.
- Test on real work. A hello world prompt tells you nothing about an agent model.
The context window is large, but a large window does not remove the need to check results. Always read what the model gives you before using it with customers or clients.
Julian also notes that the community feedback is early. More tests will come in over the next weeks, and these will give a clearer picture than any single launch post.
07Why Ornith 1.5 122B matters for open models
Julian sees a pattern in this release. Six months ago, a 122B model with these benchmark scores would have needed serious infrastructure, with multiple GPUs and a lot of setup. Now it runs on one card and pulls on Ollama.
The hardware needs are still real. Even so, the direction is clear: powerful models get more accessible with every release cycle.
He also says most business owners have not worked out how to use open models like Ornith 1.5 in their workflows. Most people are still copying and pasting into a chat box. Nobody in their network has built a real agent pipeline yet.
Inside the AI Profit Boardroom, Julian has a 30 day roadmap for agent automation with open models. Day one sets up your first agent workflow on a real business task. Day two builds a research agent that browses, summarizes and formats information.
Day three runs a content automation pipeline that drafts, plans and schedules.