What's Changed
- Records: Kapitsa against MLX on the M5 Pro, and Flash-Next old/new builds (docs only) by @alish-nace-23 in #940
- Records: harness results and generator for the Kapitsa-against-MLX M5 Pro report (docs only) by @alish-nace-23 in #942
- docs(research): history lane and selector update measured live on the 35B (+7 % and +6 % tokens per cycle) by @alish-nace-23 in #943
- Qwen prefill on a grouped encoder (NI_PF_GROUPED): independent projections and expert GEMMs side by side, bitwise the same: 4K prefill 671.7 -> 657.9 ms on the 35B-A3B (-2.1%), Clover -2.2%, 4B -2.0%, Flash-Next -1.1% on the M5 Max by @fearampgreed in #941
- Drex Max prefill: bidirectional tensor attention (qattn_t_g6d2s8_bidi) for noncausal G = 6, a 16K request 59.4 -> 50.2 s (~15% faster) on the M5 Pro by @alish-nace-23 in #938
- DFlash2's context commit on the tensor tile (qmt_m64_n64), the same tokens: a 4K prompt's prefill and commit -1.9% on the 35B-A3B (706 -> 692.5 ms), Clover -1.8%, 4B -2.9%, 9B -3.2%, 27B -1.3% on the M5 Max by @fearampgreed in #946
- Kapitsa 0.2.28: sets the version in CMakeLists.txt, the KAPITSA_VERSION fallbacks in src/, install.sh's example and packaging/homebrew/kapitsa.rb, as #937 did for 0.2.27. The tag v0.2.28 goes on the merge commit. by @fearampgreed in #948
Full Changelog: v0.2.27...v0.2.28