22 April 2026
MLX LM v0.31.3 reports cache, tool-calling, and tokenizer fixes
According to the official release notes, MLX LM v0.31.3 reports a set of bug fixes centered on cache extension, tool calling, tokenizer behavior, and model-specific issues, plus a thread-local generation stream. The supplied evidence does not establish independent performance, compatibility, or upgrade outcomes.
MLX LM v0.31.3 is a release record dated 22 April 2026. Its first-party notes present the update primarily as a collection of bug fixes, alongside a thread-local generation stream associated with MLX v0.31.2. [1]
01
What we know now
- 01
[1] MLX LM, v0.31.3 official release record, published 2026-04-22: https://github.com/ml-explore/mlx-lm/releases/tag/v0.31.3
- 02
The release notes list fixes for cache extension, parallel tool calls, tokenizer handling, several model-specific paths, and a thread-local generation stream.
02
Why this matters for Armenia
The release may be relevant to developers in Armenia who use MLX LM in local AI development workflows, but the supplied record provides no Armenia-specific adoption or impact evidence.
03
Cache handling
Batch dimension and cache-extension fixes are listed.Tool calling
Server and model-specific tool-call handling changes are listed.Tokenizer and models
Tokenizer, embedding, parser, and model-path fixes are listed.Generation stream
A thread-local generation stream is listed as a highlight.This categorizes items reported in the first-party release notes; it does not measure their effect.
04
Release overview
The official MLX LM v0.31.3 release page was published on 22 April 2026. The source presents the release as an incremental maintenance update rather than providing benchmark results or a broad compatibility assessment. [1]
- The release notes characterize v0.31.3 as containing many bug fixes.
- They also list a thread-local generation stream as a highlight associated with MLX v0.31.2.
05
Caching and model-path changes
Several reported changes concern cache behavior, particularly batch-dimension mismatches during cache extension. The notes also identify targeted fixes in cache-related code and model-specific processing. These statements are first-party descriptions, and the packet does not include independent verification of behavior after upgrade. [1]
- BatchKVCache, BatchRotatingKVCache, and ArraysCache extension paths are included in the listed fixes.
- The notes also mention a missing cache-module import and a Gemma 4 KV-shared-layer issue involving unused projections.
06
Tool calling and tokenizer handling
The release notes report fixes across several tool-calling paths. Listed items include parallel handling on the server, MiniMax M2 behavior, Gemma 4 parsing of hyphenated function names and braces within string arguments, and Mistral behavior when a tool-call end marker is empty. Tokenizer handling is also represented through a reported TokenizerWrapper fix. [1]
- The server's parallel tool-call handling is listed as fixed.
- The notes identify MiniMax M2 parallel tool calling, Gemma 4 parsing, and Mistral tool-call termination among model-specific cases.
- TokenizerWrapper handling of a None-type check for think tokens is also listed.
07
Other reported items
The release record additionally lists fixes affecting Apertus embeddings and dwq safetensors detection. It highlights a thread-local generation stream, but the supplied material does not provide implementation details, performance data, or instructions for adopting it. [1]
- An Apertus tie_word_embeddings fix is listed.
- A dwq check for actual safetensors in a target directory is listed.
- The thread-local generation stream is the only item identified as a highlight beyond bug fixes.
08
Practical next steps
For teams considering this release, the source record supports a targeted review rather than a blanket upgrade recommendation.
- 01
Compare local use of cache extension, server-side tool calling, tokenizer wrappers, and affected model paths with the fixes listed in the release notes.
- 02
Review the linked release record and associated change references for components used in a local workflow.
- 03
Test upgrades in an appropriate non-production environment, since the supplied material includes no compatibility matrix or independent validation.
09
Limits of this edition
The available evidence is limited to the official release record.
No independent testing, performance results, compatibility matrix, or upgrade guidance was supplied.
The release notes identify changes but do not establish how each change will affect a particular deployment.
SRC
Source desk
Direct links to the material behind this selection. Seeing the source matters as much as reading the synthesis.