All pull requests

Pull requests

• Route prepared MiniMax text packs through mlx-stream

  #972·sjordan0228 opened Oct 25, 2026·5 tasks done·

  ·

• Schedule two isolated streamed DeepSeek requests

  #971·sjordan0228 opened Oct 25, 2026·5 tasks done·

  ·

• Parse native MiniMax reasoning and tool calls

  #970·sjordan0228 opened Oct 25, 2026·5 tasks done·

  ·

• Resume streamed DeepSeek conversations from SSD checkpoints

  #969·sjordan0228 opened Oct 25, 2026·5 tasks done·

  ·

• fix(loop-stop): cut a restatement loop with a fixed head and a reworded tail

  bugengine
  #968·lborloz opened Oct 25, 2026Contributor·5 tasks done·

  ·
  1

  4

• app: attach plain-text files to a chat (fixes #841)

  enhancementi18nUI
  #967·ianalitis opened Oct 25, 2026·

  ·
  1

  1

• fix: reject unusable Qwen3-TTS reference audio

  bugdocumentationengine
  #966·ianalitis opened Oct 24, 2026·

  ·
  2

• fix: extend long-context latency histogram bounds

  engineenhancement
  #964·ianalitis opened Oct 24, 2026·

  ·
  3

• feat(pull): HF_ENDPOINT points every hub download at a mirror

  engineenhancementUI
  #963·lborloz opened Oct 24, 2026Contributor·5 tasks done·

  ·
  1

  2

• fix: align sidebar agent titles and summaries

  enhancementUI
  #962·ianalitis opened Oct 24, 2026·

  ·
  1

  5

• fix: gate stream-pack vision metadata by backend

  bugengine
  #961·ianalitis opened Oct 24, 2026·

  ·
  1

  5

• fix(mtp): load MoE heads that ship one tensor per expert (#848)

  bugenginemodel-support
  #960·brandondyal opened Oct 24, 2026Contributor·

  ·
  3

• perf(scheduler): cache-aware admission behind MLX_SERVE_PREFIX_WAIT

  #955·cowboycoderhq opened Oct 23, 2026Contributor·5 tasks done·

  ·

• perf(scheduler): grammar, logprobs and penalty slots stay in the decode batch

  #954·cowboycoderhq opened Oct 23, 2026Contributor·5 tasks done·

  ·

• fix(disk-cache): an SSD-first commit keeps its target entry across the budget refresh

  bugengine
  #949·lborloz opened Oct 23, 2026Contributor·5 tasks done·

  ·
  1

  1

• fix(gemma4): read per-tensor group-size overrides in the PLE table and vision projections

  bugenginemodel-support
  #948·lborloz opened Oct 23, 2026Contributor·5 tasks done·

  ·
  1

• lib/mlx-stream at ac3b772: davidtai/mlx-stream main after #27 merged

  dependenciesengine
  #942·davidtai opened Oct 22, 2026Contributor·

  ·
  1

• tests: build_mtp_sidecar.py stacks a MoE head's experts into switch_mlp banks

  enhancementmodel-support
  #941·Dankpaws opened Oct 22, 2026Contributor·

  ·
  1

• kv-attn: faster matmul2d packed attention, and row blocks for verify widths up to 16 rows

  engineperformance
  #940·Dankpaws opened Oct 22, 2026Contributor·

  ·
  4

• dflash(qwen4): DFlash2 on Qwen3.8-Flash-Next (incoai/Qwen3.8-Flash-Next-DFlash2)

  enginemodel-supportperformance
  #935·davidtai opened Oct 21, 2026Contributor·

  ·
  5

• load: MLX_SERVE_NOCACHE_LOAD=1 reads any pack's weights past the page cache

  engineperformance
  #934·davidtai opened Oct 21, 2026Contributor·

  ·
  1

• build: bump lib/sushi to sushi v1.3.0

  dependenciesengine
  #933·beamivalice opened Oct 21, 2026Contributor·

  ·
  1

• fix(codex): launch local models with the current tool contract (kill-switched)

  bugdocumentationengineUI
  #921·kmahara opened Oct 20, 2026Contributor·

  ·
  4

• feat(dsv41): keep resumable states for several conversations (#897)

  engineenhancementperformance
  #915·tpischke opened Oct 19, 2026·

  ·
  1

  2

• Control automatic model discovery and refresh My Models

  enhancementUI
  #907·trigger2k20 opened Oct 18, 2026·5 tasks done·

  ·
  1