Search results

GLM-5.2 (glm_moe_dsa with IndexShare) fails to load — missing per-layer indexer params

Status: Open.

#1418 In ml-explore/mlx-lm;

· anjansastry-hash opened on Jun 20, 2026

DeepSeek-V4-Flash drifts from Simplified Chinese to Traditional Chinese in the mlx-vlm DeepSeek V4 path

Status: Open.

#1404 In ml-explore/mlx-lm;

· BeCreatedP opened on Jun 14, 2026

feat: batched generation for Mamba/hybrid models (Nemotron-H)

Status: Open.

#1403 In ml-explore/mlx-lm;

· HaoXuAI opened on Jun 14, 2026

fetch_nearest_cache deep-copies the cached KV, doubling peak memory exactly when a cached conversation is reused

Status: Open.

#1395 In ml-explore/mlx-lm;

· VitorLudke opened on Jun 11, 2026

Model type diffusion_gemma not supported.

Status: Open.

#1391 In ml-explore/mlx-lm;

· WSINTRA opened on Jun 11, 2026

mlx_lm.server aborts with Metal OOM after prompt cache grows to ~23-26 GB

Status: Open.

#1390 In ml-explore/mlx-lm;

· agisilaos opened on Jun 10, 2026

Feature request: support <3 bpw GGUF quantization formats in MLX / mlx-lm

Status: Open.

#1388 In ml-explore/mlx-lm;

· pearsonkyle opened on Jun 10, 2026

Gemma 3n: mlx_lm.server returns garbage output (shared-KV layers incompatible with batched cache)

Status: Open.

#1384 In ml-explore/mlx-lm;

· jonlarale opened on Jun 9, 2026

Missing support for laguna model architecture ("Model type laguna not supported")

Status: Open.

#1378 In ml-explore/mlx-lm;

· dougdaulton opened on Jun 8, 2026

mlx_lm.server assigns 36-char UUID tool-call ids, which Mistral v3 templates reject (length-9 constraint)

Status: Open.

#1375 In ml-explore/mlx-lm;

· rayanhanader opened on Jun 8, 2026

Mistral tool parser rejects the JSON-list output format, dropping valid tool calls (e.g. Ministral-8B-Instruct-2410)

Status: Open.

#1374 In ml-explore/mlx-lm;

· rayanhanader opened on Jun 8, 2026

feat: expose per-token logprobs from batch_generate()

Status: Open.

#1358 In ml-explore/mlx-lm;

· HaoXuAI opened on Jun 7, 2026