Furiosa SDK Release 2026.5.0b1

Furiosa SDK Release 2026.5.0b1#

Furiosa SDK 2026.5.0b1 is the second beta release of 2026.5.0. The changes since 2026.5.0b0 are summarized below.

All Changes#

Furiosa-LLM#

Model Support & Multimodal Serving#

  • Added BF16 support for Qwen3.5-0.8B (Qwen3_5ForConditionalGeneration) on four processing elements (PEs) of a single RNGD NPU. Its FXB can be compiled with --tensor-parallel-size=4.

  • Refactored multimodal preprocessing and caching, including processor reuse across requests.

Engine Core & Performance#

  • Added token log-probability reporting with NPU sampling.

  • Improved L3 KV-cache prefetch to reduce memory pressure and redundant transfers, and preserve cached prefixes during remote fetches.

Bug Fixes#

  • Fixed experimental NPU sampling for Solar Open2 models.