Ornith 1.5 gained a 122-billion-parameter model on October 20, 2026. The dated changelog on Ollama separates that release from the 9B, 35B-A3B and 397B versions introduced on August 19. The October milestone expands the family; it does not restart the launch date for every Ornith 1.5 model.
A new size between existing options
The Ollama listing provides a 122B tag with text and image input and a listed 256K context window. Its available packages occupy roughly 73–77 GB. That is a distribution size, not a promise that the same amount of system memory will support every context length or workload.
For teams considering controlled inference, the new option creates another checkpoint to evaluate between smaller local models and the largest family member. It does not by itself establish which option delivers the lowest operating cost. The relevant comparison includes the machine that can serve it, the tasks it completes correctly and the time people spend correcting its work.
What self-improvement means here
The Ornith research explanation describes an expanded reinforcement-learning loop: task generation, scaffold construction and solution rollouts improve together. This is a training approach. It should not be read as a claim that a downloaded model will autonomously update its own production weights while answering business requests.
The publisher positions 122B around instruction following, task decomposition and workflow understanding, and makes efficiency and competitive-performance claims. Those are publisher claims, rather than results from Nerova testing. In particular, a hardware slogan about a single NVIDIA system is not a complete procurement specification.
Serving requirements extend beyond the download
Our deployment assessment starts with the exact package and runtime. A model’s weight files share memory with the runtime, image processing, the attention cache and concurrent requests. A long advertised context does not establish acceptable latency at maximum context, especially when several users submit work together.
A useful capacity trial would therefore record the quantization, available memory, prompt length and concurrency alongside each result. Include both short instructions and realistic document or image inputs. Measure time to a usable answer, peak memory and failure behavior. These are suggested measurements, not published Ornith results.
Distribution also leaves a licensing decision. Confirm the license attached to the exact artifact and its dependencies before commercial deployment; a model being listed in a local runner is not a substitute for reviewing its terms.
Evaluate workflow completion, then tool access
For an internal assistant, begin with a read-only task such as turning a supplied procedure into a checklist. Compare the output with the original instructions and record omitted steps, unsupported additions and contradictions. A model that writes a plausible plan still needs to demonstrate that it respects the supplied constraints.
If the evaluation later includes tools, the application should enforce access rights and require approval for consequential actions. Model output should request an action; the owning system should validate and execute it. Failed writes must remain failures, regardless of how confidently the assistant describes completion.
The October release is concrete: Ornith 1.5 now has a separately dated 122B option. Its operational value depends on the chosen package, serving environment and accepted work, rather than the family name or parameter count alone.