All pull requests

Pull requests

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

  #1185·nazeshinjite opened Oct 6, 2026·

  ·

• Enable the measured Metal choices on M4

  #1184·jiiim opened Oct 5, 2026·

  ·

• 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

• Keep the disk KV checkpoint the next request loads when storing an evict snapshot

  #1170·chopchop-jiahao opened Oct 2, 2026·

  ·
  1

• Preserve exact tool-call replay across KV checkpoint restores

  #1169·mvdbos opened Oct 2, 2026·

  ·

• Back off speculative drafting on consecutive rejected proposals

  #1168·emilianbold opened Oct 2, 2026Contributor·

  ·

• Add n-gram prompt-lookup speculative drafting

  #1167·emilianbold opened Oct 1, 2026Contributor·

  ·
  3

• Server: advertise loaded model modalities

  #1165·kyuz0 opened Oct 1, 2026Contributor·

  ·
  1

• Unterminated tool call: defer to the parse-failure recovery instead of erroring the turn

  #1163·nazerim opened Oct 1, 2026·

  ·

• server: report max_output_tokens as the Responses incomplete reason

  #1162·Laurianti opened Oct 1, 2026·

  ·
  1

• server: report Anthropic stop_sequence when a client stop matches

  #1161·Laurianti opened Oct 1, 2026·

  ·
  1

• Render GLM 5.3 low reasoning effort as Low

  #1160·chopchop-jiahao opened Oct 1, 2026·

  ·

• Fix GLM imatrix collector reading stale MoE buffers

  #1158·chopchop-jiahao opened Oct 1, 2026·

  ·

• Find the cold KV anchor in ChatML prompts

  #1157·paperniuk opened Sep 30, 2026·

  ·

• Fix the iq1_s and iq4_nl block sizes in the GGUF type table

  #1156·paperniuk opened Sep 30, 2026·

  ·
  1

• Fix misleading prefill progress and false live cache hits after a backend state reset

  #1155·emilianbold opened Sep 30, 2026Contributor·

  ·
  1

  1

• Metal: overlap Qwen3.8 PLE reads with prefill

  #1154·mvdbos opened Sep 29, 2026·

  ·

• server: --image-prefix-snapshot reuses the text before a request's image (#1007)

  #1153·denovich opened Sep 29, 2026·

  ·