Nous Research发布基于Qwen3.8-27B的Arete-27B模型
Nous Research发布Arete-27B，基于阿里Qwen3.8-27B并用约5300万Token进行西方政治价值观微调。其自测称，模型宣传率由22%降至约1%，通用能力未明显下降。
Loading…
bibibi.ccbibibi.ccLoading…
Connect scattered updates into a story you can understand.
Nous Research发布Arete-27B，基于阿里Qwen3.8-27B并用约5300万Token进行西方政治价值观微调。其自测称，模型宣传率由22%降至约1%，通用能力未明显下降。
Thinking Machines发布Inkling 2 Small，Artificial Analysis综合评分达31分，Agent任务和编程测试成绩较上一代提升。
百度正式开源文心5.5 Flash Base基座模型。该模型采用MoE架构，总参数量2800亿，支持最长100万Token上下文。
百度开源2800亿参数文心 ERNIE-5.5-Flash-Base模型权重，并采用较上一代更严格的自定义许可证。
Ricursive Intelligence发布开源模型Ripple-65B Preview，称其在Terminal-Bench 2.1编程测试中得分80.5%，模型权重采用Apache 2.0许可证。
PANews称，腾讯混元发布新一代大模型Hy4，新增视觉理解和计算机使用能力，并计划在未来数日开源模型权重与系统卡。
印度AI公司Sarvam发布开源决策模型Prism，称其基于105B模型并将单次请求中位响应时间降至44毫秒。模型已按Apache 2.0协议开放权重。
腾讯混元发布Hy4正式版，称扩大训练规模并新增视觉理解和电脑操作能力，已上线腾讯云、OpenRouter等平台，权重计划近期开源。
动察 Beating AI 报道称，腾讯微信视觉团队开源 WeVisOne-VL，并推出2B和4B版本，支持检测、计数、OCR、定位、分割和精细抠图。
Beating AI breaking news reports that Xiaomi has open-sourced the Xiaomi-OCR-0 model and related code, services, and Agent Skill for document recognition, field extraction, and visual question answering.
Beating AI reports that Mistral has open-sourced improved Large 4(Le Chonk) weights, providing FP8 and NVFP4 versions. The report also discloses information on its programming, efficiency, and deployment licensing.
10000 Wu announced cooperation with Hangzhou and Linping District, receiving hundreds of millions of yuan in investment from state-owned assets at both the municipal and district levels. Its headquarters will be established in Hangzhou’s Linping District, and an industrial AI transformation base will be built.