v0.9.5

What's Changed

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

New Contributors

• @wsj040303 made their first contribution in #2461
• @mikemikimike made their first contribution in #2463

Full Changelog: v0.9.4...v0.9.5