Search results

Mistral Tool Calling: Premature termination on empty tokens

Status: Open.

#1069 In ml-explore/mlx-lm;

· vitaly-d opened on Mar 28, 2026

Is it posibble to implement Anthropic style api /v1/message ?

Status: Open.

#1068 In ml-explore/mlx-lm;

· beginor opened on Mar 28, 2026

Bug: tool_calls arguments converted to dict breaks chat templates that expect string (str + dict TypeError)

Status: Open.

#1065 In ml-explore/mlx-lm;

· mauricepruna opened on Mar 27, 2026

Qwen3.5-35B-A3B-4bit can emit malformed tool-call output around 20k prompt tokens.

Status: Open.

#1061 In ml-explore/mlx-lm;

· Christ9038 opened on Mar 26, 2026

PolarQuant KV cache compression (TurboQuant, ICLR 2026)

Status: Open.

#1060 In ml-explore/mlx-lm;

· rachittshah opened on Mar 26, 2026

Qwen3.5 4B(merged LORA adapter) convert to MLX generates incorrect results

Status: Open.

#1058 In ml-explore/mlx-lm;

· clcy1029 opened on Mar 26, 2026

[Feature]: DoLA and SLED decoding strategies

Status: Open.

#1053 In ml-explore/mlx-lm;

· gkatenin opened on Mar 25, 2026

Nemotron-Cascade-2: thinking mode infinite loop on Apple Silicon

Status: Open.

#1050 In ml-explore/mlx-lm;

· roeex5 opened on Mar 24, 2026

kimi k2.5 kvcache oom crash on mac studio m3 ultra

Status: Open.

#1047 In ml-explore/mlx-lm;

· jonipupet opened on Mar 23, 2026

Question about performance optimizing & Metal Debugger profiling result

Status: Open.

#1046 In ml-explore/mlx-lm;

· MollySophia opened on Mar 23, 2026

Add KV cache quantization support to server

Status: Open.

#1043 In ml-explore/mlx-lm;

· ivanfioravanti opened on Mar 22, 2026

Devstral (tekken v13 tokenizer) produces Ġ characters instead of spaces in output

Status: Open.

#1041 In ml-explore/mlx-lm;

· luiz290788 opened on Mar 22, 2026