What's Changed
- README: Clover 35B-A3B on the M5 Max, 6,002 / 319 tok/s with the 35B's drafter (poster row, tables, benchmarks write-up, run records; docs only) by @fearampgreed in #777
- sdk: add an in-process C ABI with bounded GPU lock waits by @chainhelen in #770
- README: supported backends with a roadmap, in place of Splash's engine by @alisher-amantay in #783
- README: Features, a supported-models table with Drex 2.0 and seven Mac sizes, What's new cut to four by @alisher-amantay in #784
- sdk: return synchronous loading and Metal init errors (#782, part 1) by @chainhelen in #788
- catalog: Clover-35B-A3B as
clover(4-bit pack on the Hub with the 35B's DFlash2 drafter):kapitsa serve --model clover, 417 tok/s on an agent turn (M5 Max) by @fearampgreed in #785 - sdk: return expert-load and deferred Metal errors (584 logit rows unchanged) by @chainhelen in #793
- Qwen 4 configs checked before the download: a new qwen4* model_type, or a qwen4_exp config outside the Flash-Next path, is refused with each difference (CPU only) by @fearampgreed in #797
- perf: reuse shared decode reduction (M4 Pro 20-split -13–15%) by @chainhelen in #798
- README: the supported-models footnote in four sentences, not four paragraphs by @alisher-amantay in #790
- Reject unsupported quantization before reading weights (#787) by @alisher-amantay in #791
- Benchmarks: Kapitsa 0.2.23 against Splash 1.4.1 on an EC2 M4 Pro (macOS 26.7), Kapitsa on seven models (15.8) by @alisher-amantay in #792
- README: Quickstart comes right after the nav row, with a limits line by @alisher-amantay in #794
- README: the benchmark poster collapses and its footer no longer clips by @alisher-amantay in #795
- README: Features lead with measured outcomes, release notices are one line by @alisher-amantay in #796
- Pi: provider maxTokens is the context window, not 25% of it; opencode and Claude Code unchanged (#803) by @alisher-amantay in #804
- packaging/splash: Kapitsa's Splash patches against pinned 1.4.0, a build script that fails on a patch that stops applying (Not verified on a GPU) by @alisher-amantay in #806
- README: show the benchmark poster expanded instead of in a collapsed section by @alisher-amantay in #805
- Agent and review rules: a new feature ships on by default for one model and Mac, or is widely established by @alisher-amantay in #801
- kapitsa serve --engine splash runs KFlash for Flash-Next Q2_0: decode 1.118x stock Splash 1.4.0 (1.167x at default size); 4K prefill 1.014x at 20G, 0.989x at default size (--splash-kflash auto|on|off) by @alisher-amantay in #807
- KFlash moves to Splash 1.5.0: decode 1.101x, 4K prefill 1.004x the Splash 1.5.0 release (score 1.051, LB 1.030; same patches, SDK 26.5 vs 27.0) by @alisher-amantay in #808
- Load f16 scales/biases and 3/4/5/6/8-bit MTP sidecars (#809). Not verified on a GPU by @alisher-amantay in #810
- Flash-Next 180B Q2 poster row re-measured with KFlash on Splash 1.5.0: 1218 / 61.3 against Splash 1.5.0's 1178 / 56.5, still a tie by @alisher-amantay in #811
- Native Drex Max v1 scoring in systemone: noncausal Qwen3.5 and a readout head. Not verified on a GPU by @alisher-amantay in #812
- Kapitsa 0.2.25: sets the version in CMakeLists.txt, the KAPITSA_VERSION fallbacks in src/, install.sh's example and packaging/homebrew/kapitsa.rb, as #778 did for 0.2.24. The tag v0.2.25 goes on the merge commit. by @fearampgreed in #786
Full Changelog: v0.2.24...v0.2.25