v0.8.23

bytedance/deer-flowv0.8.23Jun 25, 2026by EricLBuehler

AI Summary

Adds support for CUDA 13.3, Gemma4 optimizations, and Hunyuan models, along with improvements to quantization and installation handling.

Key Highlights

Support CUDA 13.3
Gemma4 optimizations (shared KV layers)
Hunyuan v1 dense and MoE support
isq executor and planning improvements

New Features

CUDA 13.3 support
Gemma4 optimizations
Hunyuan v1 dense and MoE support
isq executor
`lazy_preload` macro

Full Release Notes

## What's Changed
* feat(cuda): support CUDA 13.3 by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/2275
* chore: bump candle dep by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/2276
* feat(cli): improve --quant and --isq docs in cli by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/2277
* feat(gemma4): do not load projections for shared kv layers by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/2281
* feat(quant): add isq executor and planning by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/2283
* feat: add Hunyuan v1 dense and MoE support by @ASheng1019 in https://github.com/EricLBuehler/mistral.rs/pull/2268
* feat(install): fix handling when there are preexisting installs by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/2284
* feat(gdn): add isq support by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/2285
* Fix reversed FCFS priority in PagedAttentionScheduler preemption by @pjdurden in https://github.com/EricLBuehler/mistral.rs/pull/2250
* Validate GGUF special token ids against vocab to prevent OOB panic by @pjdurden in https://github.com/EricLBuehler/mistral.rs/pull/2282

## New Contributors
* @ASheng1019 made their first contribution in https://github.com/EricLBuehler/mistral.rs/pull/2268
* @pjdurden made their first contribution in https://github.com/EricLBuehler/mistral.rs/pull/2250

**Full Changelog**: https://github.com/EricLBuehler/mistral.rs/compare/v0.8.22...v0.8.23