Furiosa SDK 2026.3: A New Kernel Framework and the Models It Enables

FuriosaAI has announced SDK 2026.3. The core of this update is a new kernel framework called TCL (Tensor Contraction Language) and the furiosa-kernels package built on top of it.

The existing approach had a limitation in that decomposing PyTorch models into ATen operations diluted the model's intent and structure. In contrast, TCL is a declarative Python eDSL that treats Tensor Contraction operations as first-class primitives, helping the compiler optimize tiling, scheduling, fusion, and more while preserving the model's structure.

This release provides the following key features:

• Native multimodal (Vision-Language) serving support via Qwen3-VL
• Fast onboarding support for large-scale MoE (Mixture-of-Experts) models such as gpt-oss, Solar-Open, K-EXAONE, Qwen3 MoE
• Model deployment without recompilation via the FXB (Furiosa Executable Bundle) format
• Introduction of an opt-in overlap scheduler that maximizes NPU utilization

Thanks to TCL's modular design, even when new architectures emerge, existing building blocks like RMSNorm, Linear, and MLP can be reused, enabling very fast model support.