AI BriefingFuriosa SDK 2025.2: Hugging Face and Reasoning Model Support
Key point
Furiosa released SDK 2025.2 for RNGD, strengthening reasoning model support and Hugging Face integration.
Details
Furiosa has released Furiosa SDK 2025.2.0 for RNGD, its chip for LLM and agentic AI inference. This update focuses on improving developer convenience and expanding the scope of model support.
Key updates include:
- Support for Reasoning Models, including DeepSeek-R1-Distill, and reasoning parsers
- Ability to build models directly from the Hugging Face Hub in bfloat16, float32, float16 formats without a separate quantization process
- Tool calling conversation functionality via Chat with tool support in the LLM API
- Chunked prefill support for large-scale prefill processing
- New metrics endpoint added for metadata and monitoring
In particular, integration with the Hugging Face ecosystem has been strengthened, allowing model artifacts to be loaded directly via furiosa-llm serve and other tools, with optimized pre-compiled models also available for immediate use.
In addition, environment compatibility has been improved through support for Ubuntu 24.04 and Python 3.11/3.12, and the Linux setup process for NPU device access has been simplified, making development environment setup even more convenient.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.