v0.18.0

What's Changed

• site: the new design (dark, framed, grain; hero with recorded examples; interactive bench chart) by @cobanov in #123
• site: the inner pages in the new design, bench included; no analytics on /try by @cobanov in #125
• release: the desktop Windows build and the cuda12 image reuse the release's engine (COB-276, COB-281) by @cobanov in #121
• parity: --smoke, a fixed subset of the goldens, and the test tiers K1–K3 by @cobanov in #120
• site: the grain is consistent across screens and covers a phone's full height by @cobanov in #128
• site: gliclass's cost barely grows with the options on a GPU, not on a CPU by @cobanov in #129
• site: model pages with an aside and grouped tags; blog posts with a table of contents by @cobanov in #130
• site: the copy button no longer overlaps the code block's corner or shows text through it by @cobanov in #134
• site: cheaper to open on a phone; grain no longer reads as mud on a 3x screen by @cobanov in #135
• docs: coba's technical report, a first draft for arXiv by @cobanov in #133
• scripts: remote-job.sh sends a command to hope-wsl or choso-wsl in tmux and waits for it (COB-280) by @cobanov in #131
• mlx: DeBERTa-v3 on the Apple GPU, phase 3 (nli:deberta-v3-large, gliclass:large) (COB-212) by @cobanov in #132
• desktop: a desktop app's layout, with the website's design language and the gaze mark by @cobanov in #126
• train: coba's cards print the release measurements by @cobanov in #136
• train: the coba card's parity count is the release run's by @cobanov in #137
• docs: coba's report closes its 14 TODOs from the training records by @cobanov in #139
• mlx: Qwen3 and Qwen3.5 on the Apple GPU, phase 4 (qwen3guard, decider) (COB-212) by @cobanov in #140
• coba: the training data is about 400,000 questions in 344,000 states by @cobanov in #143
• mlx: Decision 1.0 and kev on the Apple GPU; several weights files, a LoRA merged at load (COB-212) by @cobanov in #142
• sura:2b and sura:0.8b, Ollaya's own decoder decision models (COB-552) by @cobanov in #138
• decider: 2b v12 and 4b v3 from Mapika's 2026-10-20 update, decider-ai 1.10.0 as the reference, noul temperature 0.05 (COB-559) by @cobanov in #144
• d1-omni: Liquid AI's d1-omni-600M, text path, as a new ONNX family (COB-563) by @cobanov in #145
• d3: Decision 3.0 (edge, lite, nano, mini, flash) on the pplx engine, layout d3-code-readout-v1 (COB-558) by @cobanov in #146
• COB-560: gliclass:v4-edge, decima:base-clean, winnow:e2b, jeb v2.1 by @cobanov in #147
• strands: Gemma 4 E2B and E4B tags (strands-decider:e2b, :e4b) with layout strands-pointer-v2 (COB-564) by @cobanov in #149
• COB-562: jeeves:9b on Jeeves v2 by @cobanov in #148
• gliclass-v4: Knowledgator's GLiClass v4 2B and 9B (Qwen3.5 decoder-KV) as the new family gliclass-kv (COB-564) by @cobanov in #150
• decima: ordinal choices and the decima-guard-web preset (COB-494, #80) by @cobanov in #151

Full Changelog: v0.17.0...v0.18.0