Issues

• Tag cadence — 3.32.3 (2026-09-30), ~95 commits on main since

  ml-explore/mlx-swift-lm#797 · Filter by author markfietje opened on Nov 13, 2026
  1

• Run Build and Test on pushes to main

  ml-explore/mlx-swift-lm#791 · Filter by author aleroot opened on Nov 11, 2026

• Next Release (3.34.0?)

  ml-explore/mlx-swift-lm#788 · Filter by author davidkoski opened on Nov 9, 2026
  2

• Add batched generation

  ml-explore/mlx-swift-lm#771 · Filter by author ronaldmannak opened on Nov 6, 2026
  2

• [BUG] Guided generation is not repeatable because schema field order follows dictionary hashing

  ml-explore/mlx-swift-lm#768 · Filter by author alexmodrono opened on Nov 5, 2026
  1

• [BUG] Generation traps after an MLX error inside withError

  ml-explore/mlx-swift-lm#762 · Filter by author mattt opened on Nov 2, 2026

• [BUG] granite-docling-258M-mlx stops the process with Idefics3Processor

  ml-explore/mlx-swift-lm#759 · Filter by author mattt opened on Nov 2, 2026

• [BUG] GLM-OCR throws for images with a side under 28 pixels

  ml-explore/mlx-swift-lm#758 · Filter by author mattt opened on Nov 2, 2026

• Qwen3.5/3.6 MTP verify runs the Gated DeltaNet recurrence twice; it can be one pass

  ml-explore/mlx-swift-lm#753 · Filter by author Augustas11 opened on Nov 2, 2026

• Qwen2VL: vision attention spans all images in a request, mixing multi-image answers

  ml-explore/mlx-swift-lm#752 · Filter by author Augustas11 opened on Nov 2, 2026

• [Feature Request] Add native support for Liquid AI LFM3-1.7B and LFM3-VL-2B

  ml-explore/mlx-swift-lm#741 · Filter by author alvarogarciapiz opened on Oct 28, 2026

• Proposal: LiquidAI d1 decision models (d1-3B, d1-omni-600M)

  ml-explore/mlx-swift-lm#694 · Filter by author ykhrustalev opened on Oct 13, 2026
  1

• Qwen3-VL vision attention ~3.5x slower on M5: head dim padded to 80 misses the NAX SDPA kernel in vendored MLX v0.32.2

  ml-explore/mlx-swift-lm#691 · Filter by author yk-vfc opened on Oct 13, 2026

• MLXEmbedders: add ModernBERT (Granite Embedding R2) support

  ml-explore/mlx-swift-lm#686 · Filter by author MBark286 opened on Oct 8, 2026
  3

• A shared core for model code & and multimodal input for embedders

  ml-explore/mlx-swift-lm#682 · Filter by author CodebyCR opened on Oct 7, 2026
  9

• MLXRerankers: Qwen3 name check admits rerankers with a different scoring protocol; JinaForRanking ignores sliding-window layers (jina-reranker-v3.5)

  ml-explore/mlx-swift-lm#673 · Filter by author magicnight opened on Oct 4, 2026
  5

• ChatSession re-reads the whole prompt every turn on hybrid models (Qwen3.5 and other MambaCache models)

  ml-explore/mlx-swift-lm#660 · Filter by author jhancock1975 opened on Oct 1, 2026

• Detect GLM-4-9B-0414 tool calls that have no markers

  ml-explore/mlx-swift-lm#658 · Filter by author thechriswebb opened on Sep 30, 2026
  1

• Consider leveraging MLXLMCommon or refactor box code

  ml-explore/mlx-swift-lm#655 · Filter by author thechriswebb opened on Sep 29, 2026

• Next Release (3.33.0)

  release-planning
  ml-explore/mlx-swift-lm#648 · Filter by author davidkoski opened on Sep 28, 2026
  12

• Expose a public transactional MTP step API for cached/batched callers

  ml-explore/mlx-swift-lm#645 · Filter by author Augustas11 opened on Sep 28, 2026

• RFC: Branchable KV cache with shared-prefix / copy-on-write semantics

  ml-explore/mlx-swift-lm#629 · Filter by author slamfinger opened on Sep 18, 2026
  5

• Read Gemma tool-call arguments with a dialect parser instead of rewriting them into JSON

  ml-explore/mlx-swift-lm#628 · Filter by author Adi2K opened on Sep 17, 2026

• RotatingKVCache cannot trim, so a serving layer cannot rewind a divergent prefix

  ml-explore/mlx-swift-lm#618 · Filter by author AtlantisPleb opened on Sep 11, 2026
  4

• DFlash2 speculative decoding for Qwen3.8-27B

  ml-explore/mlx-swift-lm#606 · Filter by author spokvulcan opened on Sep 4, 2026
  1