v0.9.5

EricLBuehler/mistral.rsv0.9.5Oct 19, 2026by EricLBuehler

Upgrading?

AI Summary

This release introduces support for the Qwen3.8-Flash-Next model, updates CUDA compatibility to version 13.4, and makes Metal Paged Attention the default backend. It also includes fixes for tool-call delimiters, GCC 13+ compatibility, and quantization deadlock issues.

Key Highlights

  • Added support for the Qwen3.8-Flash-Next model
  • Updated CUDA support to version 13.4
  • Metal Paged Attention is now the default backend
  • Fixed deadlock issues in quantization batching

New Features

  • Support for Qwen3.8-Flash-Next model
  • Support for CUDA 13.4
  • Metal Paged Attention as default backend

Full Release Notes

## What's Changed
* fix(core): only honor tool-call and reasoning delimiters emitted as special tokens by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/2455
* feat(models): add support for Qwen3.8-Flash-Next by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/2462
* feat(cuda): support CUDA 13.4 by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/2481
* fix(cuda): support GCC 13+ in radix top-k kernel by @wsj040303 in https://github.com/EricLBuehler/mistral.rs/pull/2461
* fix(quant): avoid deadlock while batching ISQ output jobs by @mikemikimike in https://github.com/EricLBuehler/mistral.rs/pull/2463
* feat(metal): make Metal Paged Attention the default backend by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/2493
* chore: v0.9.4 -> v0.9.5 by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/2498

## New Contributors
* @wsj040303 made their first contribution in https://github.com/EricLBuehler/mistral.rs/pull/2461
* @mikemikimike made their first contribution in https://github.com/EricLBuehler/mistral.rs/pull/2463

**Full Changelog**: https://github.com/EricLBuehler/mistral.rs/compare/v0.9.4...v0.9.5

Related Repositories