大模型
www.aibase.com ·
deepin 25 适配进迭时空 K3 芯片，在 RISC-V 平台实现离线大模型本地推理
据 AIbase 报道，deepin 25 在 deepin-ports SIG 支持下完成对进迭时空（SpacemiT）新一代 RISC-V 芯片 K3 的深度适配，通过底层工具链升级和本地 AI 运行时集成，可在小硬件平台上运行本地 AI 模型'小悟'，并称这是 RISC-V 平台桌面操作系统首次具备开箱即用的离线大模型推理能力。K3 采用 RVA23 规范与异构融合架构，集成 8 个 X100 通用核（最高 2.4GHz、约 130 KDMIPS）、8 个 A100 AI 核（自研 IME 矩阵扩展、约 60 TOPS）、2 个 RT24 实时核及 3D GPU、视频编码引擎，典型功耗 15-25W。deepin 25 保留 6.18 内核分支，并引入 Next 上层叠加仓库，将工具链升级至 GCC 15/16、Binutils 2.46、Glibc 2.42、LLVM 21/22、OpenSSL 3.5。上述性能与'首次'表述来自转述，尚无官方或独立实测验证。
AI推荐理由 输入提供了 deepin 25 适配进迭时空 K3（SpacemiT K3）RISC-V 芯片的具体技术细节：RVA23 规范、8 核 X100 通用核（2.4GHz、约 130 KDMIPS）、8 核 A100 AI 核（自研 IME 指令集、约 60 TOPS）、2 核 RT24 实时核、15-25W 典型功耗，以及工具链升级到 GCC 15/16、Binutils 2.46、Glibc 2.42、LLVM 21/22、OpenSSL 3.5 等。这些是较具体的新信息，对 RISC-V 桌面生态和端侧离线大模型推理有实际参考价值。局限在于：来源为 AIbase 转述，非 deepin 或进迭时空官方公告，正文未给出可验证的实测数据或发布链接；'30B 模型 10+ tokens/s'、'首次实现开箱即用离线大模型推理'等表述缺乏独立证据，需保留不确定性。
72
正文
deepin 25 Successfully Adapts to Jindie Space K3 Chip, Achieving Offline Large Model Local Inference on RISC-V Platform
AIbase基地
Published inAI News · 6 min read · Sep 17, 2026
9
In the development of open source and RISC-V ecosystem, domestic operating systems have achieved a key technical breakthrough. Recently, with the full support of the deepin-ports SIG (Special Interest Group), Deepin Operating Systemdeepin 25 has successfully completed a deep adaptation to the next-generation flagship RISC-V chip SpacemiT K3 by Jindie Space. With the upgrade of the underlying toolchain and the integration of the local AI runtime, the system not only smoothly runs the local AI model "Xiao Wu" on small hardware platforms, but also marks the first time that the desktop operating system on the RISC-V platform truly possesses the capability for offline large model inference out-of-the-box. As an RVA23 specification computing platform specifically designed for the AI era, Jindie Space K3 adopts advanced RISC-V heterogeneous fusion computing technology. Inside a single chip, it cleverly integrates three complementary core types, building an efficient "general + AI + real-time" three-domain collaborative architecture. Among them, 8 X100 general-purpose computing cores have a maximum clock speed of 2.4 GHz and about 130 KDMIPS of overall general computing power, responsible for handling system control and daily desktop interactions; 8 A100 AI computing cores specially designed for AI workloads are equipped with a self-developed matrix operation extension (IME) instruction set, providing about 60 TOPS of native AI computing power, and can maintain a smooth inference speed of more than 10 tokens per second when running a 30 billion parameter large model. In addition, the chip is equipped with 2 RT24 real-time cores for system management, a 3D GPU, and a video encoding engine. With a typical power consumption of 15 to 25 watts, it balances high energy efficiency and rich application scalability. To make the new hardware perform at its best, deepin 25 has made comprehensive deep optimizations in the system architecture. In addition to maintaining the 6.18 kernel branch, u-boot, and dedicated GPU drivers, the development team has innovatively introduced the Next repository as an upper-layer overlay repository. This repository fully enables RVA23 global optimization for the riscv64 architecture, upgrading the core toolchain to natively support the RVA23 Profile GCC 15/16, Binutils 2.46, Glibc 2.42, LLVM 21/22, and OpenSSL 3.5 with various algorithm assembly-level optimizations. This higher baseline basic runtime environment not only lays a solid foundation for fully tapping into the hardware instruction potential of the upper software, but also accumulates valuable technical verification experience for deepin's subsequent cross-major version migration. In the highly anticipated local AI capabilities construction, this adaptation has successfully打通 the complete chain from the underlying toolchain to the application layer. The built-in Xiao Wu student deeply integrates with the RVA23 and SpacemiT matrix operation extension through deepin-modelhub, and can automatically schedule the optimal operator according to the chip architecture. At the same time, the widely used GGML computing library for mainstream model runtime such as llama.cpp has fully opened support for SpacemiT-related instruction sets, achieving complete vector and matrix computation acceleration. Test results show that when users run local AI models in the system, the system can intelligently dispatch computing tasks to the 8 A100 AI cores for efficient processing. Currently, the latest deepin 25 RISC-V experience image adapted to K3 has been officially released in the deepin-ports image list. Developers and open source enthusiasts can download and complete the flashing through official channels. With the continuous improvement of the Next repository's basic construction, this breakthrough not only allows RISC-V desktop devices to truly have the ability to run large models offline, but also opens up broad space for the collaborative innovation of domestic open source software and hardware ecosystems.
RISC-VdeepinSpacemiTK3AI
This article is from AIbase Daily
Scan to view
Welcome to the [AI Daily] column! This is your daily guide to exploring the world of artificial intelligence. Every day, we present you with hot topics in the AI field, focusing on developers, helping you understand technical trends, and learning about innovative AI product applications.
—— Created by the AIbase Daily Team
© Copyright AIbase Base 2024, Click to View Source -