All pull requests

Pull requests

• slide: /v1/slide/learn takes questions beside its text

  #992·philip-pentatonic opened Oct 17, 2026Contributor·8 of 9 tasks·

  ·
  3

• kolibri1: Kolibri-1 on CUDA through the lane host, drafts off (#481)

  #991·jschmied opened Oct 17, 2026Contributor·8 of 9 tasks·

  ·

• RUNBOOK: what the heat bands did on three GB10s, GLM-5.3-Flash NVFP4 at TP=3 with a 1M-token prompt

  #988·Bizuayeu opened Oct 17, 2026Contributor·8 tasks done·

  ·

• music3: MiniMax Music 3 lyrics to song on Metal, from Zig (stacked on #982)

  #987·drowzeys opened Oct 17, 2026·

  ·

• Flash Next: images and video on native Metal, and warm passes that keep nothing skipped (stacked on #606)

  #986·BobClawblaw opened Oct 17, 2026Contributor·

  ·
  1

• drama: backbone attention runs the reference's 64-key softmax tiles

  #985·plotarmordev opened Oct 17, 2026Contributor·4 of 9 tasks·

  ·

• GLM-5.3-Flash on two CUDA ranks: 16 streams at a 40960-token window

  #984·BobClawblaw opened Oct 17, 2026Contributor·

  ·

• music3: MiniMax Music 3 lyrics to song on one CUDA GPU, from Zig (new family; proposal)

  #982·drowzeys opened Oct 17, 2026·

  ·

• benchmark: attached runs read a server manifest, so a two-rank CUDA run can be ranked

  #980·tovereast opened Oct 17, 2026·

  ·

• GLM-5.3-Flash on two CUDA ranks: DFlash2 drafts (--drafter)

  #979·BobClawblaw opened Oct 17, 2026Contributor·

  ·

• Qwen3.5/3.8 on CUDA: decode 37-73% faster on #836's engine, bits unchanged (on #836)

  #978·BobClawblaw opened Oct 17, 2026Contributor·8 of 9 tasks·

  ·

• Gemma 4 at every MLX affine width, uniform or per module: oQ packs on the native Metal engine (stacked on #975)

  #977·BobClawblaw opened Oct 16, 2026Contributor·

  ·

• GLM-5.3-Flash on two CUDA ranks: video input

  #976·BobClawblaw opened Oct 16, 2026Contributor·

  ·
  1

• Gemma 4 12B on the native Metal engine: text, images and its MTP assistant (stacked on #879)

  #975·BobClawblaw opened Oct 16, 2026Contributor·

  ·
  2

• fix(template): append in place when a namespace string accumulates

  #973·nmcbride opened Oct 16, 2026·

  ·

• qwen3_5_moe on CUDA: 3-bit routed experts beside affine-4 (on #856)

  #972·CanadaApollo6 opened Oct 16, 2026·8 tasks done·

  ·
  2

• fix: skip macOS AppleDouble files in the model library scan

  #970·heitke opened Oct 16, 2026Contributor·

  ·

• fix: let the model library scan read large checkpoint metadata

  #967·heitke opened Oct 16, 2026Contributor·

  ·

• fix(flashnext): a HIP round's graph replays only over its pool slot's caches

  #963·BlivionIaG opened Oct 16, 2026Contributor·9 tasks done·

  ·

• MiniMax H3 on Metal: Ref2VA, video and audio from reference pictures, videos and sound

  #961·drowzeys opened Oct 16, 2026·

  ·
  3

• MiniMax H3 on Metal: the released MiniMax-H3 weights, and re-voicing a finished clip

  #960·drowzeys opened Oct 16, 2026·

  ·
  1

• MiniMax H3 / FastH3 on Metal: any step count, with the schedule and AdaLN tables made natively

  #959·drowzeys opened Oct 16, 2026·

  ·
  1

• MiniMax H3 / FastH3 on CUDA in the checkpoint's bf16, one GPU or several Sparks over RoCE

  #958·drowzeys opened Oct 16, 2026·

  ·
  3

• docs: describe Metal checks in the macOS test suite

  #957·cwschroeder opened Oct 16, 2026Contributor·

  ·

• drama3: Fish Audio Drama 3 text to speech on Metal, from Zig (stacked on #939)

  #956·drowzeys opened Oct 16, 2026·

  ·
  7