Issues

• Support Ternary Bonsai 2 27B (MLX 2-bit) with reasoning and vision

  lmstudio-ai/mlx-engine#377 · Filter by author vokasug opened on Oct 2, 2026

• Bump bundled MLX VLM 0.6.5 to 0.7.1 to add muse_glimmer model support

  lmstudio-ai/mlx-engine#376 · Filter by author andrzejsorowka-ux opened on Sep 28, 2026
  1

• BatchedModelKit does not respect multiple EOS token IDs from model config

  lmstudio-ai/mlx-engine#374 · Filter by author iphizic opened on Sep 5, 2026

• MLX server (#353) as an Engine Protocol runtime: use case + offer to help

  lmstudio-ai/mlx-engine#373 · Filter by author archmagos-certiv opened on Sep 2, 2026
  5

• glm5_next (GLM-5.3-Flash): prompt cache save always fails — save_safetensors on zero-size array in batched-vision blob store; plus latent hybrid-cache bug in CacheWrapper._num_tokens_in_cache

  lmstudio-ai/mlx-engine#372 · Filter by author tobieapb opened on Aug 31, 2026

• Draft-only (MTP) repos indexed as loadable models, then crash at generation: 'Qwen3_5MTPDraftModel' object has no attribute 'get_input_embeddings'

  lmstudio-ai/mlx-engine#371 · Filter by author kafasam opened on Aug 30, 2026
  1

• [Feature] MLX engine ignores built-in MTP (nextn) heads — native MTP speculative decoding missing on the MLX path

  lmstudio-ai/mlx-engine#370 · Filter by author nikolai-vysotskyi opened on Aug 21, 2026
  1

• Batched vision models cannot use a draft model: is_draft_model_compatible() returns False unconditionally

  lmstudio-ai/mlx-engine#368 · Filter by author cristyanul opened on Aug 19, 2026

• MLX auto-fit replaces the configured context length and disagrees with the prompt cache budget

  lmstudio-ai/mlx-engine#366 · Filter by author philippdubach opened on Aug 15, 2026
  2

• Partial GPU Offloading support for MLX Qwen3.6-35b-a3b 8bit on M4 Mac 48GB?

  lmstudio-ai/mlx-engine#365 · Filter by author combacsa opened on Aug 13, 2026

• Backend scheduler crashes with broadcast_shapes ValueError in BatchRotatingKVCache.merge under concurrent requests with shared prefix

  lmstudio-ai/mlx-engine#363 · Filter by author seb760 opened on Aug 7, 2026

• repeat_penalty causes fatal backend-scheduler crashes in multi-turn conversations (32-bit slice error / KV cache merge broadcast error)

  lmstudio-ai/mlx-engine#359 · Filter by author thefrenchbaron opened on Aug 5, 2026

• Text-only architectures implemented only in mlx_vlm (e.g. deepseek_v4) fail to load because dispatch checks only for vision_config

  lmstudio-ai/mlx-engine#357 · Filter by author True2456 opened on Aug 1, 2026

• Disk prefix KV caching

  lmstudio-ai/mlx-engine#354 · Filter by author richardwerkman opened on Jul 24, 2026
  1

• Please support for Ternary Bonsai 27b with vision an reasoning

  lmstudio-ai/mlx-engine#347 · Filter by author lustfeind opened on Jul 15, 2026
  4

• Bug: Batched VLM loader treats mtp.safetensors as base weights for Qwen3.5-4B-OptiQ-4bit

  lmstudio-ai/mlx-engine#343 · Filter by author Zrxrxrx opened on Jul 8, 2026
  3

• Feature: fused batched decode for concurrent requests (close the gap to mlx_lm.batch_generate)

  lmstudio-ai/mlx-engine#342 · Filter by author nikolai-vysotskyi opened on Jun 30, 2026

• ​[Feature Request] Add configuration toggle to disable prompt disk caching to prevent excessive SSD wear

  lmstudio-ai/mlx-engine#341 · Filter by author houzy opened on Jun 25, 2026

• We hope to upgrade the mlx-vlm version to 0.6.2 or above to resolve the issue of VLM lacking the qwen3_5_moe_vision type, which prevents the Holo 3.1 model from being used.

  lmstudio-ai/mlx-engine#339 · Filter by author MuShan-bit opened on Jun 16, 2026
  3

• Gemma 4 (26B-A4B) reasoning never terminates on MLX — fills max_tokens with reasoning_content, empty content (GGUF terminates fine)

  lmstudio-ai/mlx-engine#337 · Filter by author DocTpoint opened on Jun 14, 2026

• DiffusionGemma

  lmstudio-ai/mlx-engine#336 · Filter by author AlejandroJL opened on Jun 12, 2026
  1

• Qwen3.6 35B A3B Mxfp8 - failed to load

  lmstudio-ai/mlx-engine#330 · Filter by author sharki13 opened on May 24, 2026
  2

• kv cache starts back from 0 every other turn or so in all models and in all harnesses

  lmstudio-ai/mlx-engine#327 · Filter by author coolaj86 opened on May 15, 2026
  1

• Image tokens not included when running Qwen3.6-35B-A3B MLX (tokens: 0, features N)

  lmstudio-ai/mlx-engine#325 · Filter by author ripple7511 opened on May 11, 2026
  1

• Support for model type Zaya?

  lmstudio-ai/mlx-engine#324 · Filter by author Phatcat opened on May 10, 2026