DecisionTune 1.0: a 395M offline decision model picking options in ~10ms on Mac
Abe238 · reddit · 2026-10-05
A developer released DecisionTune 1.0, a 395M local decision model (ModernBERT-large + 4KB scoring head, Apache-2.0). Given a state, question and option list, it returns per-option probabilities or P(yes) in a single encoder pass — no text generation.
- Performance: median 9.6ms per short decision on an M5 Pro Mac with MLX (1.7GB GPU memory); 65ms CPU-only; 1.58GB fp32 weights, 8,192-token context cap.
- Backends: PyTorch / MLX / ONNX agree on 99.85% of 2,755 questions; 50 parity rows verified across all three before each release.
- Offline: no network after first download; SHA-256 manifest checks.
- Quality & limits: Decision Index 0.2.1 score of 29.57, strongest at Tools & Automation (46.5); option-picking only, uncalibrated probabilities, English only, weak at knowledge/math/taste.
Ships with a CLI (uvx decision-tune ask), browser app, and an MCP server so local assistants can offload routing and yes/no checks.
More from coding & agent
- OpenAI ships compaction in Responses API; devs debate if it kills differentiation — zeeg · 2026-10-06
- Dev builds Claude Code skill that writes better HTML plans with plain language, mockups and linting — trq212 · 2026-10-06
- OpenAI ships compaction in Responses API, sparking vendor lock-in debate among developers — pvncher · 2026-10-06
- DoorDash launches MCP and CLI for agentic ordering; dev auto-restocks office pantry with camera — Scobleizer · 2026-10-06
- zeeg warns labs against vendor lock-in via compaction: 'focus on models, not locking customers' — zeeg · 2026-10-06
- Anthropic ships Claude Code mods: TypeScript functions that rewrite prompts and UI — thione · 2026-10-06