All pull requests

Pull requests

• Add a native macOS desktop chat app

  #605·linson007 opened Oct 27, 2026Contributor·

  ·

• Local classifier for Claude Code auto mode: POST /v1/safeguards and --safeguards

  #591·sambengtson opened Oct 25, 2026·

  ·
  1

• Publish Qwen architecture ablations and DFlash2 deployment audit

  #590·linson007 opened Oct 25, 2026Contributor·

  ·

• 加入 GGUF Metal 編譯問題的選用 register 與分段解碼路徑

  #587·Ryan-cat-1489 opened Oct 25, 2026·

  ·

• Use smaller SIMD-group rows for short affine FFN prefills on M5 Pro

  #585·linson007 opened Oct 24, 2026Contributor·

  ·
  1

• Use M16 for measured one-lane AF4G64 decode on M5 Pro

  #584·linson007 opened Oct 24, 2026Contributor·

  ·

• Report complete decode cycles and publish Qwen M5 Pro evidence

  #583·linson007 opened Oct 24, 2026Contributor·

  ·
  1

• Add Qwen3.5-4B target and DFlash2 draft support

  #559·Qiqi723 opened Oct 23, 2026·

  ·

• Add matched-prefix DFlash state restore probes and native evidence

  #556·linson007 opened Oct 22, 2026Contributor·

  ·
  1

• Add matched-context DFlash proposal policy overlap reference

  #555·linson007 opened Oct 22, 2026Contributor·

  ·
  1

• Add selective recall training and reached-prefix coverage metrics

  #554·linson007 opened Oct 22, 2026Contributor·

  ·
  1

• Add training-only pair-context and nonlinear selector prototypes

  #553·linson007 opened Oct 22, 2026Contributor·

  ·
  1

• Add experimental prefix-survival loss and position weighting

  #552·linson007 opened Oct 22, 2026Contributor·

  ·
  1

• Add experimental loader-matched AF4G64 projection training

  #551·linson007 opened Oct 22, 2026Contributor·

  ·
  1

• Add DFlash acceptance capture and offline selector analysis

  #550·linson007 opened Oct 22, 2026Contributor·

  ·
  1

  1

• Add ability to use custom templates

  #549·Ptico opened Oct 22, 2026·

  ·
  2

• Add experimental Apple Silicon MPS draft training

  #537·linson007 opened Oct 21, 2026Contributor·

  ·
  1

• Serve a model without a draft with --draft-model none

  #534·Flo5k5 opened Oct 21, 2026·

  ·
  1

  1

• Steer Qwen3.8-Flash-Next, Qwen3.8-27B and Qwen3.6-35B-A3B with a control vector (--control-vector)

  #532·Flo5k5 opened Oct 21, 2026·

  ·
  1

• Encode chat-template tags quoted in message text as plain text

  #530·Flo5k5 opened Oct 21, 2026·

  ·

• Add Qwen serving benchmarks for verification policy and mixed workloads

  #523·linson007 opened Oct 20, 2026Contributor·

  ·
  1

• Add --model-dir for local target model directories

  #505·trigger2k20 opened Oct 20, 2026·

  ·
  1

• Run a one-lane MoE step's few-row experts through a register GEMV (opt-in --moe-numerics fast)

  #503·zaiyang-zhang opened Oct 20, 2026Contributor·

  ·

• Serve MiniMax-H3 video as a sibling serve mode (M5 Pro)

  #502·zaiyang-zhang opened Oct 19, 2026Contributor·

  ·

• Add a live engine performance dashboard at /dashboard

  #501·vjanelle opened Oct 19, 2026·

  ·