-
1
英伟达出70亿收编Poolside
英伟达向AI初创Poolside投资10亿美元并支付60亿美元技术授权费,逾百名员工转入英伟达负责Nemotron项目
Nvidia Poolside deal to compete with Chinese Open Weights ↗3.6 -
2
Kimi K3推理成本190美元/M
开发者实测2.8万亿参数Kimi K3,8卡B300部署下解码92 tok/s,推理成本190美元/百万token
I hosted Kimi K3 (2.8T parameters) using 8 B300s. 92 tok/s, $190 per million tokens ↗3.4 -
3
Qwen3.8-27B被赞编码超Luna
多条独立帖称Qwen 3.8 27B编码、OCR表现比肩甚至超过GPT Luna,社区口碑持续发酵
Qwen 3.8 27B is a game changer. ↗3.0 -
4
DeepSeek瘦身跑上M2 Ultra
开发者将DeepSeek V4 Flash无损压缩至141GiB(比Q4 GGUF还小),在M2 Ultra上跑出25.8-42 t/s
DeepSeek V4 Flash on an M2 Ultra: repacked to 141 GiB losslessly, smaller than the Q4 GGUF, at 25.8 t/s (42 t/s peak) ↗3.0 -
5
AMD分支让llama.cpp提速2倍
AMD维护的llama.cpp分支在Strix Halo等设备上实现预填充速度最高2倍提升,已持续合并进上游
AMD Users: Have you tried the llamma.cpp AMD-Ecosystem branch? Up to 2x PP Speed ↗2.3 -
6
玩家扩容至36台DGX Spark集群
个人家庭实验室将DGX Spark集群从16台扩容至36台,配200Gbps网络,统一内存达4.6TB
“The All Spark” Cluster: Upgrading from 16 - 36 DGX Sparks ↗2.1 -
7
开源工具让Claude Code喊你
开源项目Agent Quest新增功能,在Claude Code或Codex会话需要人工介入时发出视觉与声音提醒
Agent Quest now tells you when Claude Code or Codex needs you visually and with sound ↗2.1 -
8
GMKtec将推Ryzen AI新机
GMKtec计划在IFA Berlin 2026发布搭载AMD Ryzen AI Max+ PRO 495的新款迷你主机
GMKtec is going to launch new hardware with Ryzen AI Max+ PRO 495 at IFA Berlin 2026 ↗2.0 -
9
开发者自制LLM文本水印工具
开发者受Anthropic计划为模型输出加水印启发,实现了一个SynthID-Text风格的教学向LLM水印方案
Implementing Watermarking for Language Models [P] ↗2.0