mistral.rs (Eric Buehler’s Rust LLM inference engine, not affiliated with Mistral AI) merged PR #2462 on 2026-10-01: feat(models): add support for Qwen3.8-Flash-Next.
This is a model-support merge, not a new release tag, and it does not repeat the 0.9.3 / 0.9.4 coverage. The root LICENSE is MIT.
What landed
The PR adds Qwen/Qwen3.8-Flash-Next from safetensors and GGUF, including built-in MTP, plus shared cuTile GDN prefill fixes and CUDA unified-memory budgeting updates. Hard limits per the PR: paged attention requires CUDA for this model; FP8 KV caches, tensor parallelism, GGUF MTP, and external drafters are unsupported (#2462).
Author-reported GB10 HTTP benches vs llama.cpp (same UD-Q4_K_XL, BF16 KV, 16k context) claim prefill ~55–58% faster with decode essentially tied: attribute; not independently verified (#2462).
Related: hipfire v0.4.0 is the AMD/RDNA counterpart story for Flash-Next, covered in its own post.
Who should care
mistral.rs users who need Qwen3.8-Flash-Next on CUDA should read #2462 and keep the unsupported-feature list in the runbook.
Update, 20 October 2026: now in a tagged release, v0.9.5
The Qwen3.8-Flash-Next support from PR #2462 now ships in a tag. mistral.rs v0.9.5, published 19 October, includes it alongside CUDA 13.4 support (#2481), Metal Paged Attention as the default backend on Apple GPUs (#2493), and a fix so tool-call and reasoning delimiters only count when emitted as special tokens (#2455). As with earlier tags, crates.io still lists mistralrs 0.8.1, so treat the GitHub tag as the version source.

The Campfire
CommentsThanks. Your comment is waiting for the desk. Once it is approved it appears here within about a minute.
Add a comment