Apodex 1.1 mini

A 36B mixture-of-experts model from Apodex AI, post-trained for long-horizon agentic work: planning, calling tools, running code and recovering from failures inside one continuous task. It is a fine-tune of Qwen3.5-35B-A3B and inherits that architecture — 256 experts with 8 active per token, so roughly 3B parameters do the work on any given step, and a hybrid attention stack where only every fourth layer is full attention. That keeps the KV cache small enough that long contexts stay affordable on a single card. Community GGUF builds put a 4-bit quantisation inside 24 GB; the vendor's own NVFP4 and Int4 checkpoints are larger and belong to the 32 GB tier. Apodex publishes its agent runtime, FrontierAgent, under the same Apache-2.0 licence. All benchmark figures on the model card were produced by that runtime and scored by the vendor.