Colibrì's Brio mode scores candidate answers instead of generating text locally
Just_Vugg_PolyMCP · reddit · 2026-09-21
Colibrì founder Vincenzo announced v1.12.0 with a new "Brio mode" inspired by TypeSafe: instead of generating a reply, it scores a set of allowed answers, returning a probability for each option plus the entropy of the distribution.
- Reuses models Colibrì already supports locally — no new model to download or train
- Shared context is reused across questions via snapshots, with a different candidate set allowed per question
- The dashboard shows probability bars per option and entropy per question; no output tokens are generated, though the engine still processes context and scores option tokens
- The author stresses these are normalized scores over your options, not calibrated confidence — low entropy doesn't mean the answer is right
Accessible via dashboard, terminal, or POST /v1/brio; the repo is open source.
More from Infra
- Daniel Lemire Tests Whether CPUs Can Take More Than One Branch Per Cycle — lemire · 2026-09-21
- Qwen3.8-27B in native 8-bit hits 37-55 tok/s on Apple Silicon, avoiding the 4-bit reasoning cliff — SnooPredictions515 · 2026-09-21
- Google's Orphaned VMs Patches Keep VMs Running While Host Kernel Goes Offline — jedisct1 · 2026-09-21
- djev-run Deploys DiffusionGemma on Cloud Run's RTX PRO 6000 Blackwell for ~$3/hr — bodonoghue85 · 2026-09-21
- 16GB (often 12GB) is the realistic VRAM ceiling for most people running local AI — ECrispy · 2026-09-21
- IDC: global AI compute gap to hit $381B by 2030 as stacking GPUs stops working — 量子位 · 2026-09-21