Ornith 1.5 35B A3B is a sparse mixture-of-experts multimodal model that activates only a fraction of its parameters per token, combining a 35B-total expert pool with 3B activated per forward pass. It is built on the Qwen3.5 MoE architecture using the Qwen3_5MoeForConditionalGeneration model class, a design choice that gives it a hybrid text-and-vision backbone while keeping inference costs closer to a small dense model. The same documentation notes support for text, image, and video understanding, alongside chat, reasoning, and tool-use abilities, framing the release as a self-improving variant aimed at coding agents that need to read screenshots, diagrams, or short clips and then act through tool calls. Bilingual coverage of English and Chinese broadens its practical reach for multilingual codebases and documentation workflows.

Open-weight availability and a 262,144-token context window make the model well suited to long-horizon agent tasks such as multi-file refactors, repository-scale code review, and tool-mediated debugging sessions where retaining earlier file contents and tool traces matters. Compatible serving engines include vLLM, Transformers, and SGLang, so it can drop into existing inference stacks with a PyTorch checkpoint and standard launch tooling. The combination of reasoning, structured output, and tool calling makes it a practical fit for teams building autonomous coding assistants or research workflows that need a controllable, locally deployable multimodal model without paying for a much larger dense frontier system.