Trans-N Ships Swama 2.5.1 With OpenAI Decisions API Support
Trans-N is positioning local decision processing, not text generation, as the practical use case, and its own test numbers put a locally hosted Qwen3.5-35B model ahead of a cloud API call on time while landing in the same accuracy range.
Reporting from 1 source: ASCII.jp.
Trans-N released version 2.5.1 of Swama, an open-source platform for running large language models locally on Apple silicon Macs. The build supports the OpenAI Decisions API format, so classification, yes/no decisions, and graded evaluations on text and images run on the machine without a cloud inference call. In internal testing on a Mac Studio (M3 Ultra), 30 inquiries classified in 4.5 seconds and 30 product photos assessed in 10.5 seconds.
The release centers on a decision endpoint, "/v1/decisions", built for routing inquiries to a department or judging whether a product photo shows damage, rather than for producing long text. Applications call it through the API and receive the result with a confidence value from the model. Trans-N notes limits on supported input formats and image counts, and states that confidence does not guarantee a correct answer; the supported scope is listed in the GitHub documentation.
The comparison numbers come with stated caveats. Both sides started at once and processed one item at a time, timed as the median of three runs through a common local proxy. Trans-N says results vary by model, input, hardware, and network conditions, and that the measurement does not demonstrate superiority in speed or accuracy across all environments.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.