Issues

• --max-cache-disk argument value not considered

  incoai/splash#446 · Filter by author baek54321 opened on Oct 17, 2026
  1

• Opt-in INT8-activation prefill for MLX affine 4/5-bit on Apple10 (1.3–1.9x prefill GEMM)

  incoai/splash#444 · Filter by author jma24 opened on Oct 16, 2026

• Decode about 2-3.5% slower per drafted token on MLX 4-bit targets since they load as block images (c5abaa4)

  incoai/splash#441 · Filter by author ezoushen opened on Oct 15, 2026

• test-performance-real: every round lacks short samples on 1.3.2 (the parser expects rows + 1 prompt tokens)

  incoai/splash#440 · Filter by author ezoushen opened on Oct 15, 2026

• test-real: the runtime oracle's first penalty probe fails on 1.3.2 without shader validation

  incoai/splash#437 · Filter by author ezoushen opened on Oct 15, 2026

• dflash2 generation repo

  incoai/splash#435 · Filter by author linson007 opened on Oct 14, 2026

• Incompressible context?

  incoai/splash#432 · Filter by author robert-werner opened on Oct 14, 2026

• Qwen3.8-Flash-Next MTP decode: the hyper-connection norm and full-vocabulary drafting take a large share of each cycle

  incoai/splash#422 · Filter by author Vardaan97 opened on Oct 14, 2026

• Hybrid models: an agent harness editing a tool result deep in a long session finds no retained state, so a 174K prompt that shares about 105K re-prefills from zero

  incoai/splash#387 · Filter by author Vardaan97 opened on Oct 12, 2026

• Stream Qwen3.8-Flash-Next's PLE n-gram table from SSD (opt-in) so larger quants fit on 128 GB Macs

  incoai/splash#372 · Filter by author Vardaan97 opened on Oct 11, 2026
  2

• Several Splash servers on one Mac run 3–7× slower when they overlap: a cross-process GPU lease?

  incoai/splash#370 · Filter by author janhilgard opened on Oct 11, 2026
  2

• [Showcase] SplashBar — a tiny macOS menu bar controller for the local Splash server

  incoai/splash#363 · Filter by author MarcelKuesters opened on Oct 10, 2026
  2

• M5 Max optimizations in place from https://github.com/publicExcess/splish/?

  incoai/splash#362 · Filter by author tomchk opened on Oct 10, 2026

• First sibling-turn cache miss: proactive composite-state checkpoint proposal

  incoai/splash#361 · Filter by author ardnt opened on Oct 10, 2026
  2

• love the work.. keep it going

  incoai/splash#360 · Filter by author 100firescarlos-cell opened on Oct 10, 2026
  3

• Proposed UI to see engine status (working link attached!)

  incoai/splash#333 · Filter by author badvision opened on Oct 7, 2026
  1

• Neural Engine split: a calibration taken on a busy Mac is remembered and keeps a smaller share (0.35 vs 0.56), ~20% slower prefill until forgotten by hand

  incoai/splash#332 · Filter by author komarov-dc opened on Oct 7, 2026
  2

• Apple9 (M4 Max 40c): sub-8-bit GGUF decode kernels reach 300-390 GB/s vs 460 for Q8_0, which caps a Q8-quality mixed GGUF at -9% per cycle

  incoai/splash#331 · Filter by author komarov-dc opened on Oct 7, 2026
  2

• Neural Engine split prefill: the ANE idles during each layer's mixer (~47% busy on M4 Max); overlapping it could add ~1.3x

  incoai/splash#328 · Filter by author komarov-dc opened on Oct 6, 2026

• Infinite paragraph-level repetition loops

  incoai/splash#326 · Filter by author hello-sebastian opened on Oct 6, 2026
  1

• Live verify rows: skip expensive work for low-confidence draft guesses

  incoai/splash#280 · Filter by author linson007 opened on Oct 3, 2026

• Decode cycle stage profiling and draft acceptance analysis

  incoai/splash#258 · Filter by author linson007 opened on Oct 2, 2026
  1

• [Showcase] Splash on iPhone 18 Pro Max: Qwen3.6-35B-A3B with experts streamed from flash (unofficial port)

  incoai/splash#250 · Filter by author dalnk opened on Oct 2, 2026

• [Feature Request] Add Qwen3.5-4B and Qwen3.5-9B models support

  incoai/splash#236 · Filter by author nailarch opened on Sep 30, 2026
  1

• Just wanted to say thanks

  incoai/splash#228 · Filter by author knopkem opened on Sep 29, 2026
  1