What's Changed
- README: M5 Pro 27B Kapitsa cell 649 / 81 -> 606 / 77 after #371 and #381 (Kapitsa-only, median of three); mlx-serve's 78 now the row's fastest decode by @alisher-amantay in #413
- Benchmark: Mac mini (2024) · M4 Pro · 64 GB · Qwen3.8 27B 4-bit by @bobosola in #420
- KFlash wraps every decode; prompt-lookup rows are opt-in (NI_KFLASH_BUDGET=1) until a cost model shows they pay by @alisher-amantay in #396
- README: plain KFlash on every Kapitsa cell (held until a7185aee's KFlash-wrapper PR merges) by @alisher-amantay in #390
- 35B-A3B MoE prefill on the M5: expert GEMM outputs staged as whole rows, +7.1% at 4K / +6.5% at 16K, bitwise identical by @alisher-amantay in #412
- Flint: the decode merges in dispatches of their own on an Ultra (M5 Ultra garbage output, #421); not verified on a GPU by @fearampgreed in #423
- Terminal background query waits with select: macOS's poll on /dev/tty returned at once, so OSC 11 / DA1 replies leaked into the shell (#415) by @fearampgreed in #424
- Releases ship kapitsa-systemone and kapitsa-drex-pointer: tarball, install.sh, Homebrew and kapitsa update (built and tested on a Mac; build.sh and brew not run) by @alisher-amantay in #418
- systemone: one encoder per prefill chunk, a render depth bound, the context's last position and recoverable ANE splits (#416, not verified on a GPU) by @alisher-amantay in #422
- Releases fail when the Homebrew tap cannot be updated: a stable release either updates the tap or goes red by @alisher-amantay in #425
- systemone: reject over-deep JSON before it reaches the parser; fix the stale live test by @alisher-amantay in #429
- Benchmark: MacBook Pro (16-inch, Nov 2023) · M3 Max · 96 GB · Qwen3.8 27B 4-bit by @AlexW68 in #430
- Benchmark: Mac Studio (2026) · M5 Ultra · 256 GB · Qwen3.8 27B 4-bit by @l0westbob in #431
- Benchmark: Mac Studio (2026) · M5 Ultra · 256 GB · Qwen3.8 27B 4-bit by @l0westbob in #432
- kapitsa doctor: the expert store a Flash-Next Q2_0 pack will get, and why (one reason function with the engine's load log) by @alisher-amantay in #433
- Benchmark: MacBook Pro (13-inch, M2, 2022) · M2 · 16 GB · Qwen3.5 4B 8-bit by @alisher-amantay in #436
- bench: name KFlash and prompt-lookup decoding, record KFlash's drafter (#295), not verified on a GPU by @alisher-amantay in #435
- Kapitsa 0.2.15 by @alisher-amantay in #437
- bench/community/RESULTS.md: rewritten from the results (69 runs), not verified on a GPU by @alisher-amantay in #438
- engine: the split chunk's wait is NI_ANE_SPLIT_WAIT_MS, and a test gives a chunk up (refs #426) by @alisher-amantay in #434
- bench: Flash-Next streamed from the SSD on the M5 Pro 48 GB, Kapitsa against llama.cpp by @alisher-amantay in #442
- Tool arguments keep each value's spelling (true, null, 2.50, [1,2], non-ASCII) through the Qwen3.5/3.6 and Nemotron templates by @alisher-amantay in #439
- Bench: Flint M5 Pro context sweep 1K-128K, Kapitsa 71d6891 vs llama.cpp b11840 (prefill 1.28-2.02x, decode 1.6x) by @alisher-amantay in #444
- logit_dump and logit_floor.py for Qwen3.8-Flash-Next, and its M5 Pro floor (bitwise on every source) by @alisher-amantay in #446
- qwen35: a dense MLP with gate and up in different formats is refused at load, not on the first request (refs #445, not GPU-verified) by @alisher-amantay in #449
- engine: a model of another architecture is refused by its model_type, not a JSON exception (not GPU-verified) by @alisher-amantay in #448
- Split: streamed Macs get layers by memory share (5.1 -> 11.0 tok/s, M3 + 24 GB M4 Pro); a half-copied model file is named, not a JSON crash by @fearampgreed in #451
- Benchmark: Mac Studio (2026) · M5 Ultra · 256 GB · Qwen3.8 Flash-Next 180B-A7B 4-bit by @l0westbob in #452
- Several Macs: groups pair when together they hold Flash-Next, a later Mac joins a split and the head re-places layers when a token gets faster; fixes #451's header check breaking test_server --cli (not verified between Macs) by @fearampgreed in #454
- docs: a split that streams places layers by memory share, not by bandwidth alone (#451) by @alisher-amantay in #456
- README: a star-history chart above the footer by @itomatik in #457
- History prefill waits for a quiet server and reads the conversation back from the SSD, so a second question after a tool loop hits the cache (two conversations, side requests) by @alisher-amantay in #455
- bench/community/RESULTS.md: the community results, rewritten from bench/community/results (70 runs, the M5 Ultra Flash-Next row) by @alisher-amantay in #453
- Benchmark: Mac Studio (2025) · M3 Ultra · 512 GB · Qwen3.8 Flash-Next 180B-A7B 4-bit by @MythicSorcerer in #458
- Benchmark: Mac Studio (2025) · M3 Ultra · 512 GB · Qwen3.8 27B 4-bit by @MythicSorcerer in #459
- Benchmark: Mac Studio (2025) · M3 Ultra · 512 GB · Qwen3.5 4B 4-bit by @MythicSorcerer in #460
- bench/community/RESULTS.md: the community results, rewritten from bench/community/results (72 runs) by @alisher-amantay in #461
- Benchmark: MacBook Pro (14-inch, 2024) · M4 Max · 64 GB · Recursive Flint 40B-A2.5B 4-bit by @shivswami in #465
- Benchmark: MacBook Pro (14-inch, 2024) · M4 Max · 64 GB · Recursive Flint 40B-A2.5B 4-bit by @shivswami in #464
- Benchmark: Mac Studio (2026) · M5 Ultra · 256 GB · Qwen3.5 4B 4-bit by @l0westbob in #468
- Benchmark: Mac Studio (2026) · M5 Ultra · 256 GB · Qwen3.5 9B 4-bit by @l0westbob in #469
- Benchmark: Mac Studio (2026) · M5 Ultra · 256 GB · Qwen3.6 35B-A3B 4-bit by @l0westbob in #470
- Benchmark: Mac Studio (2026) · M5 Ultra · 256 GB · Recursive Flint 40B-A2.5B 4-bit by @l0westbob in #471
- bench/community/RESULTS.md: the community results, rewritten from bench/community/results (79 runs) by @alisher-amantay in #466
- History prefill takes a slot when its own was used by another conversation, so --slots 1 hits the next question too by @alisher-amantay in #463
- Benchmark: Mac Studio (2026) · M5 Ultra · 256 GB · Nemotron 3.5 Bolt 74B-A9B NVIDIA 24 GB by @l0westbob in #474
- Benchmark: MacBook Pro (16-inch, 2021) · M1 Max · 32 GB · Qwen3.5 4B 4-bit by @brechtm in #476
- Flint verify passes of 5-8 rows take moe4_* too: 4.22 -> 3.99 ms/token at 8 rows, bitwise identical by @alisher-amantay in #443
- bench/community/RESULTS.md: the community results, rewritten from bench/community/results (81 runs) by @alisher-amantay in #475
- bench/community/RESULTS.md: 81 rows for 81 JSONs (the #476 row) by @alisher-amantay in #479
- Flint splits the decode kernels' merges before Apple9 for every export: an M1 Max's 4-bit one answered garbage (#478) — not verified on a GPU by @alisher-amantay in #481
- docs/usage.md: the NI_FLINT_SPLIT row (the decode merges' force, undocumented since #423) by @alisher-amantay in #483
- Kapitsa 0.2.16 by @alisher-amantay in #482
New Contributors
- @AlexW68 made their first contribution in #430
- @l0westbob made their first contribution in #431
- @MythicSorcerer made their first contribution in #458
- @shivswami made their first contribution in #465
- @brechtm made their first contribution in #476
Full Changelog: v0.2.14...v0.2.16