All pull requests

Pull requests

• Add MiMo V2.6 Flash MOPD support on Metal

  #1208·mvid opened Oct 11, 2026·

  ·

• Qwen3.8 Flash Next: +40% decode on one DGX Spark (prose), +46% (code)

  #1207·matrixfede opened Oct 11, 2026·

  ·

• server: preserve Qwen Responses final-answer cache reuse

  #1206·mvdbos opened Oct 10, 2026·

  ·

• Rewind the live GLM-5.3 KV to a snapshot instead of rebuilding the prompt

  #1205·williamxie1989 opened Oct 10, 2026·

  ·
  1

• Metal: fuse the V4.1 shared expert's gate, up and SwiGLU

  #1204·yijunyu opened Oct 10, 2026·

  ·

• server: live prefill/decode status via GET /v1/status + macOS menu bar

  #1203·williamxie1989 opened Oct 10, 2026·4 tasks done·

  ·

• Metal: retain bounded Qwen Responses prompt checkpoints

  #1200·mvdbos opened Oct 9, 2026·

  ·

• Metal: reuse Q4_K raw loads in Qwen NAX prefill

  #1198·mvdbos opened Oct 9, 2026·

  ·
  1

• Accept WebP images on macOS via ImageIO

  #1197·joohooo opened Oct 9, 2026·

  ·
  2

• responses: heal the forked live checkpoint on a token-mismatch miss

  #1194·isgallagher opened Oct 7, 2026·

  ·

• rocm: fix host-side rsqrtf() call that breaks the ROCm build

  #1193·guessleej opened Oct 7, 2026·

  ·
  1

• metal: use shuffle network for non-causal top-k block sort

  #1192·emilianbold opened Oct 7, 2026Contributor·

  ·

• Add qwen35moe CPU support for Qwen3.6 and Ornith

  #1191·simoltos opened Oct 7, 2026·

  ·

• Fix RTX PRO 6000 Blackwell CUDA builds by gating native FP4 MMQfix(cuda): resolve R6K sm_120 MMQ compilation failures

  #1190·gaowayne opened Oct 7, 2026·2 tasks done·

  ·
  1

• Add a Strix Halo (ROCm/gfx1151) benchmark harness and recorded baseline

  #1187·timlawrenz opened Oct 6, 2026·

  ·

• Qwen3.8 Flash Next exact MTP: draft from the sampled token, defer the residual, allow depth 3

  #1185·nazeshinjite opened Oct 6, 2026·

  ·
  2

• Enable the measured Metal choices on M4

  #1184·jiiim opened Oct 5, 2026·

  ·
  3

• Add Kolibri-1 support

  #1183·aparusel opened Oct 5, 2026·

  ·

• agent: persist Qwen3.8 Flash Next image sessions on Metal

  #1182·Mingke77777 opened Oct 5, 2026·

  ·

• CUDA: V4.1 expert hit profiler and resident fast path for SSD streaming

  #1180·cdome94 opened Oct 5, 2026·

  ·
  1

• Speed up Qwen3.8 Flash Next on Metal

  #1179·kernelpool opened Oct 4, 2026·

  ·
  1

• metal: run the GLM-5.3-Flash prefill router on the tuned BF16 tensor-unit matmul

  #1177·williamxie1989 opened Oct 3, 2026·

  ·
  1

• server: honor rewritten history in live tool continuations

  #1175·mvdbos opened Oct 3, 2026·

  ·

• metal: stream GLM 5.3 Flash Q2 through the expert cache, for decode and prefill

  #1174·daaain opened Oct 3, 2026·

  ·
  2

• fix(metal): cache GLM mixed Q2/Q4 routed experts

  #1172·leivy-dev opened Oct 2, 2026·

  ·
  1