Ollama 0.40.7 Release Candidate Adds MLX Inference Support for PPLX Decider v1.2

Ollama’s 0.40.7-rc0 release on October 18 adds Perplexity’s PPLX Decider v1.2 to the MLX execution backend. This gives local deployments another way to run a specialized decision model, suitable for evaluating tasks such as classification, routing, and scoring. At the time of review, GitHub still marked the version as a pre-release, so it should not be considered a full stable release. Release notes

The technical focus is the model’s execution semantics. The merged implementation covers non-causal attention, publisher-defined decision readout, image processing, and System One prompt encoding. These components must work together to use the model’s decision weights correctly; simply loading the backbone and using a standard text-generation workflow is not enough to reproduce its behavior. Merged pull request

According to the official model card, v1.2 places shared state and multiple questions in a single sequence, then reads out results at marker positions following each question’s choices. Full-attention layers can access the entire request, while linear-attention layers retain causal behavior. As a result, multiple judgments about the same input can share computation, but questions can also affect one another. The official tests show that grouping related questions can help some tasks, while scoring competing tool candidates together may reduce accuracy. Official model card

The Ollama model page now includes an example call to /v1/systemone, and the model listing offers BF16, MXFP8, and NVFP4 MLX versions, with file sizes of approximately 52 GB, 29 GB, and 19 GB, respectively. These figures indicate download size only; deployments still need to reserve runtime memory, so they should not be treated as minimum hardware requirements. Model interface, model tags

For engineering teams, the next step is to compare choice probabilities, rankings, and latency for fixed requests across the reference implementation and the MLX backend, testing single-question and multi-question formats separately. The release materials provide no comparison of MLX performance or quantized accuracy, so they do not establish the degree of acceleration. Before using the model for agent routing, teams should also check whether probability shifts invalidate existing decision thresholds.