The Pulse
Fluid Inference Converts Three Decision 3.0 Models for Apple Silicon
Fluid Inference converted three Decision 3.0 models to Apple’s Core ML format and tested them on an M5 Pro MacBook Pro. The post reports JevBench scores, request times and demos for text, screenshots and live video.
AI.info Team · Written with AI ·
Decision 3.0’s d3-mini cut median JevBench request time from 438 milliseconds to 169 milliseconds in a test on an Apple M5 Pro, according to a post by Alex Weng of Fluid Inference, published October 11, 2026. Fluid Inference converted three smaller models from the multimodal decision family—d3-lite, d3-nano and d3-mini—to Apple’s Core ML format, and built demos for text, screenshots and live video.
The release shifts those models’ decision-making onto Apple silicon. A request can be handled without a network round trip, and the post says the pull request, web page or video being assessed stays on the Mac. The Core ML models and a Swift runtime are available through the linked Hugging Face repository.
Three model sizes, tested on one M5 Pro
Fluid Inference compared the Core ML versions with upstream PyTorch on a MacBook Pro with an M5 Pro chip and 24 GB of memory. On the public JevBench suite of 231 tasks, d3-lite scored 74.0% in Core ML versus 73.2% in PyTorch; d3-nano scored 75.8% versus 75.3%; and d3-mini scored 81.8% in both.
On the 2,000-question TEST split of typed-decisions, the Core ML results were 51.6% for d3-lite, 62.7% for d3-nano and 73.9% for d3-mini. The corresponding PyTorch results were 51.6%, 62.7% and 73.6%. The post cautions that typed-decisions labels come from annotators who sometimes disagree, and that JevBench scores are exact-label results on the public suite—not the sealed leaderboard or its composite score.
Median JevBench request times fell from 204 to 53 milliseconds for d3-lite, 173 to 69 milliseconds for d3-nano, and 438 to 169 milliseconds for d3-mini. Those timings include a local HTTP round trip, and both backends ran one request at a time. Fluid Inference reports download sizes of 3.7 GB, 10.5 GB and 19.6 GB, respectively.
Core ML handles the model; the host prepares each decision
The models use a Qwen3.5 backbone and a small FP32 readout that scores the available options at the final prompt token. Core ML runs the backbone and vision tower; the host application handles token embeddings, position calculations and the 255-way readout. Each question takes one forward pass, and the model returns scores rather than generating a text answer.
At FP16 precision, the d3-mini text graph takes 6.7 GB for each length bucket. Loading two buckets at once caused GPU memory paging on the tested 24 GB Mac, with steps taking 8 to 12 seconds; the post says 8-bit weights keep both resident. The conversion used PyTorch 2.7 and coremltools 9.0.
Demos show where small decision models fit
In one demonstration, d3-lite classified 1,000 sample pull requests by vLLM-SR workgroup and change type, processing each in 0.12 seconds. On real project pull requests, d3-mini matched maintainers’ workgroup labels on 134 of 168 items and change-type tags on 153 of 182.
Another demo uses d3-mini to choose links from screenshots while navigating a book catalog. It reached all 17 goals in the test, at about 2.2 seconds per click; the post says the model compared values less reliably when asked to choose across a whole page, reaching 9 of 17 goals in that setup. For live video, the model checks the most recent four seconds—eight frames at two frames per second—and takes 1.1 seconds per check on Core ML, compared with 2.4 seconds for PyTorch on the same Mac.
The reported measurements cover one M5 Pro, GPU execution and single-request tests. They do not establish how the models perform on other Apple chips or under concurrent workloads.