Furiosa SDK Release 2026.5.0b2

Furiosa SDK Release 2026.5.0b2#

Furiosa SDK 2026.5.0b2 is the third beta release of 2026.5.0. The changes since 2026.5.0b1 are summarized below.

Added image inputs for speculative decoding, including EXAONE-4.5 MTP. Multimodal draft-model decoding requires the same FXB for target and draft; MTP remains unavailable for Qwen vision-language models.
Added guided reasoning and tool-call output for Qwen3.5, Qwen3.6, and Qwen3.8 in Chat Completions and the Responses API.

Reduced host-to-NPU data transfers for Qwen3-VL by moving text-embedding lookup to the NPU, enabling input prefetch with overlap scheduling and NPU sampling.
Improved profile-guided decode cost estimation by reducing measurement noise and timing spikes.
Refactored KV-cache memory planning.

Enabled NPU sampling by default for generative models. Use --no-enable-npu-sampling or Python SchedulerConfig(enable_npu_sampling=False) for host sampling, including when using an FXB without sampling kernels. Speculative decoding disables NPU sampling automatically; the previous --experimental-enable-npu-sampling flag remains accepted.
Changed multimodal processor options to use modality-keyed objects. For example, replace {"min_pixels": 1024} with {"image": {"min_pixels": 1024}} in the --mm-processor-kwargs CLI argument, Python LLM(mm_processor_kwargs=...), and request-level mm_processor_kwargs. Request options override defaults within each modality.
Added optional vLLM-style requantization for Mistral NeMo static per-tensor FP8 checkpoints through --experimental-enable-fp8-max-scale-requant and Python LLM(fp8_max_scale_requant=True). This matches vLLM’s weight handling at a small precision cost and is disabled by default.

Fixed unnecessary multimodal cache misses when equivalent processor options are supplied in a different order.
Fixed trailing replacement characters (U+FFFD) when generation ends with an incomplete character.