Decision models gain vision: open d3 runs free on your Mac
Five Apache-2.0 models answer choice, yes/no and score questions on text, images and video — an 8B tops last gen's 27B, and Core ML builds skip the API.
Copy markdown
Open decision models grow eyes
vLLM Semantic Router shipped Decision 3.0 — six Apache-2.0 models from 0.6B to 26B that take text, images and video and return a calibrated probability for choice, yes/no or score questions instead of generated prose. That makes them drop-in classifiers, routers and content guardrails you self-host for free, with no model output to parse.
An 8B that beats last gen's 27B
d3-flash (8.4B) scores 59.0 on the team's index, past Decision 2.0's 27B at 55.9; the flagship d3 edges Perplexity's Decider v1.1 on text (64.1 vs 62.8) and vision (71.6 vs 70.6), and hits 97.7 on CORD+FUNSD document extraction. The numbers are self-reported, so benchmark your own task — but free document-field extraction with confidence scores is now a weekend build.
Runs on your Mac, no per-call bill
Fluid Inference converted d3-lite (0.9B), nano (2B) and mini (4B) to Core ML with a Swift runtime (FluidInference/d3-coreml, FluidUse on GitHub), so decisions run locally on Apple Silicon — nothing leaves the machine and every call is free. Their demo sorts 1,000 pull requests into workgroups on an M5 Pro. Build this weekend: a local PR/issue router or an offline image-moderation filter.
Gemini CLI stops silently breaking your diffs
Gemini CLI v0.63.0 is 15 fixes, no new features. The headliner: it was overriding Git's diff.external with an empty string, the cause of 'cannot spawn' errors during diffs; it also fixes a negative truncateToolOutputThreshold that roughly doubled tool output instead of disabling truncation. Update if your diffs or output have been misbehaving.