Issues

• Proposal: ternary weights ({-1,0,1}, per-group scale) as a quantization mode

  ml-explore/mlx-lm#1966 · Filter by author Jakevin opened on Oct 12, 2026
  4

• LoRA on Qwen3.5 / Qwen3-Next: step time scales with the number of trained GatedDeltaNet layers (backward runs the per-token ops path)

  ml-explore/mlx-lm#1956 · Filter by author timothyzhbw-jpg opened on Oct 8, 2026
  1

• gemma4: convert fails after transformers 5.17 saves the config again

  ml-explore/mlx-lm#1949 · Filter by author Avicennasis opened on Oct 6, 2026

• ernie4_5_moe: e_score_correction_bias is dropped, so expert selection differs from the reference

  ml-explore/mlx-lm#1947 · Filter by author Knightkolla opened on Oct 5, 2026

• Server drops message.tool_calls while reporting finish_reason="tool_calls" (v0.32.0)

  ml-explore/mlx-lm#1946 · Filter by author santanu20 opened on Oct 4, 2026
  3

• Per-module quantization overrides are silently dropped when a model's sanitize() renames weights

  ml-explore/mlx-lm#1924 · Filter by author sje397 opened on Sep 26, 2026
  2

• server: uncaught exception in extract_cache kills the generation thread while the HTTP listener keeps accepting — all subsequent requests hang forever with no error

  ml-explore/mlx-lm#1909 · Filter by author lumatic2 opened on Sep 21, 2026
  1

• qwen3_coder tool parser returns raw string for parameters whose schema has no top-level "type" (oneOf/anyOf)

  ml-explore/mlx-lm#1907 · Filter by author petriematt opened on Sep 21, 2026
  1

• Missing general Knowledge Distillation paradigm

  ml-explore/mlx-lm#1900 · Filter by author diogomsmiranda opened on Sep 17, 2026

• mlx_lm.server: option to restrict which models a request may load (allow-list)

  ml-explore/mlx-lm#1892 · Filter by author mrcushen-arch opened on Sep 15, 2026

• RuntimeError: There is no Stream(gpu, 3) in current thread at KVCache __del__ teardown

  ml-explore/mlx-lm#1888 · Filter by author dahai80 opened on Sep 14, 2026

• Lfm2MoeSparseMoeBlock routing diverges from HF: softmax instead of sigmoid, expert_bias in the weights, and routed_scaling_factor dropped

  ml-explore/mlx-lm#1883 · Filter by author hijohnnylin opened on Sep 13, 2026
  1

• Support for k2_horizon Request.

  ml-explore/mlx-lm#1876 · Filter by author Geramy opened on Sep 10, 2026
  5

• Missing support for Orion model architecture

  ml-explore/mlx-lm#1874 · Filter by author aspwow opened on Sep 10, 2026
  1

• ArraysCache metadata graph grows during sustained Qwen3.8-27B batch decoding

  ml-explore/mlx-lm#1871 · Filter by author marksmeltzer opened on Sep 10, 2026

• Paged KV storage and external cache integration

  ml-explore/mlx-lm#1849 · Filter by author IndenScale opened on Sep 5, 2026
  1

• ArraysCache.advance() still leaks one Metal buffer per SSM layer per token through BatchGenerator after #1632 — long completions crash at ~499000/(#SSM−1) tokens on main

  ml-explore/mlx-lm#1845 · Filter by author aidiffuser opened on Sep 4, 2026

• RFC: Streaming conversion - quantize checkpoints larger than local disk in mlx_lm.convert

  ml-explore/mlx-lm#1842 · Filter by author azamamirza opened on Sep 4, 2026

• mlx_lm.server deadlocks in mlx::core::eval_impl (condition_variable::wait) when the prompt cache is full under concurrent load — no Python exception

  ml-explore/mlx-lm#1834 · Filter by author wmikeseiler opened on Sep 3, 2026

• Long-running mlx_lm.server: memory grows unboundedly past prompt-cache cap on hybrid-attention models (Qwen3.8-27B 4-bit)

  ml-explore/mlx-lm#1807 · Filter by author kafasam opened on Aug 30, 2026

• mlx_lm.server: per-request chat_template_kwargs bypasses the prompt cache (0.31.3)

  ml-explore/mlx-lm#1803 · Filter by author jackhwarner opened on Aug 29, 2026
  1

• mlx_lm.server: streamed generations do not store prompt-cache entries (0.31.3)

  ml-explore/mlx-lm#1802 · Filter by author jackhwarner opened on Aug 29, 2026
  1

• Missing support for meituan-longcat/LongCat-Flash-Lite-Sparse

  ml-explore/mlx-lm#1794 · Filter by author MattBash17 opened on Aug 27, 2026

• feat: MXFP8 / fp8 mixed-precision training path (LoRA + full finetune)

  ml-explore/mlx-lm#1793 · Filter by author dahai80 opened on Aug 27, 2026

• gated_delta: vectorized Metal kernel reads dt_bias per-head; callers (kimi_linear) pass per-element [H, head_dim] — GPU/CPU gate mismatch

  ml-explore/mlx-lm#1792 · Filter by author aidiffuser opened on Aug 27, 2026