All pull requests

Pull requests

• Edge0: run Edge0-120B and Edge0-145B (latent MoE on the Flash Next shell), with MTP (initial implementation)

  #694·Turee opened Nov 13, 2026·

  ·

• Show live decode progress on the dashboard for non-streaming requests

  #689·jvmenen opened Nov 9, 2026Contributor·

  ·

• Stop a cancelled /v1/completions request instead of decoding to max_tokens

  #687·SuperMarioYL opened Nov 7, 2026Contributor·

  ·
  1

• server: translate Responses input_image parts to chat image_url

  #686·LivenseAI opened Nov 7, 2026·

  ·

• Update transformers requirement from !=5.13.0,<5.17,>=5.10.0 to >=5.10.0,!=5.13.0,<5.23

  dependenciespython
  #684·dependabot[bot] opened Nov 7, 2026Bot·

  ·

• hc verify read: serve the M5 Ultra (byte-identical, 7.5 to 8.8% faster decode)

  #682·nomishbhardwaj opened Nov 7, 2026Contributor·

  ·

• Bump mlx from 0.32.2 to 0.32.4

  dependenciespython
  #667·dependabot[bot] opened Oct 24, 2026Bot·

  ·

• Bump filelock from 3.29.0 to 4.2.1

  dependenciespython:uv
  #663·dependabot[bot] opened Oct 24, 2026Bot·

  ·

• Bump pillow from 12.3.0 to 13.0.0

  dependenciespython:uv
  #662·dependabot[bot] opened Oct 24, 2026Bot·

  ·

• Faster MoE routing on M1

  #660·a3rosol opened Oct 23, 2026Contributor·

  ·

• Add mtplx serve --jit: load models on demand

  #659·jregeimbal opened Oct 23, 2026·

  ·

• SSD cold tier: the streaming spill uses the incremental-encode memo too

  #651·jvmenen opened Oct 20, 2026Contributor·

  ·

• SSD cold tier: read restores in parallel into one buffer per tensor

  #648·jvmenen opened Oct 19, 2026Contributor·

  ·

• SSD session cache telemetry: where restores spend their time, where they land, and how much the cache writes

  #647·jvmenen opened Oct 19, 2026Contributor·

  ·

• SSD cold tier: evict and expire entries this server can never restore first

  #646·jvmenen opened Oct 19, 2026Contributor·

  ·

• GDN prefill on MLX's chunked gated_delta_update (MTPLX_GDN_FAST_PREFILL, off by default)

  #642·jvmenen opened Oct 19, 2026Contributor·

  ·

• Segmented KV: the segment kernel follows the quarter split and the wide verify route

  #639·jvmenen opened Oct 17, 2026Contributor·

  ·

• Add MiniCPM-V 4.7 adapter and forward canvas attention arguments

  #633·shiftedx opened Oct 13, 2026Contributor·

  ·

• Segmented KV on Macs without tensor units (M1-M4): per-segment attention with a logsumexp merge

  #630·jvmenen opened Oct 13, 2026Contributor·

  ·

• Cut per-turn host cost of prefix lookups (cold-tier scan, prompt conversion, RAM compares)

  #629·jvmenen opened Oct 13, 2026Contributor·

  ·

• Share one common_prefix_len and one reader per session-prefix setting

  #628·jvmenen opened Oct 13, 2026Contributor·

  ·

• Run the fixed-M4 stage-3 lane for q2/g64 routed experts

  #625·trustos opened Oct 12, 2026·

  ·

• Dashboard: show what the server really runs with on the Settings tab

  #623·jvmenen opened Oct 11, 2026Contributor·

  ·

• Keep an unwritten idle conversation through a WARNING pressure trim

  #620·jvmenen opened Oct 10, 2026Contributor·

  ·

• Write an idle conversation to SSD before the admission releases it

  #618·jvmenen opened Oct 10, 2026Contributor·

  ·