大模型前沿速递

• MiniMax raises ~160B HKD through 95.41B HKD share placement and 65B HKD convertibles, directing 80% (~127.66B HKD) to AI infrastructure and model...

• GLM-5.2 drives 40% of OpenRouter tokens, overtaking US models in weekly volume while offering 72-82% lower pricing for agentic coding.
• DeepSeek-V4...

• 核心融资信息：趋境科技完成A轮融资，半年内累计超10亿元，由河南投资集团汇融基金领投，老股东超额跟投。
• 投资方背景：领投方正联合推进日均数万亿级Token工厂建设，依托绿色能源与算力基础设施资源。
• 赛道战略意义：高品质AI Token从单点验证进入规模化供给，单位算力产出成关键竞争力，支撑企业生产级落地。

Two complementary approaches enhance LLM reasoning and task execution.

• Reasoning Expert Graph via InfraNodus structures knowledge visually, bridges...

Frontier agentic AI in hospitals introduces dynamic risks from autonomy and multi-system interactions that static ethics frameworks cannot handle.
-...

SambaNova刚完成10亿美元F轮融资，估值达110亿美元，较今年2月E轮后估值翻4.5倍。

• 估值飙升主因：2月推出SN50芯片，宣称推理速度是Blackwell B200的5倍以上，支持10万亿参数模型。
• 关键客户落地：摩根大通选其搭建私有化推理底座，凸显头部金融机构自建安全算力趋势。
• 战略定位：与英特尔深化异构数据中心合作，SN50 2026下半年交付，软银为首家伙伴。

Modulate ranks #1 on Hugging Face’s Open ASR Leaderboard out of 88 models, excelling in real-world noisy audio via 500M training hours and...

Over 200 researchers and economists including Jack Clark, Jeff Dean and Yoshua Bengio released a July 13 statement calling on governments to prepare...

World models target physical simulation rather than language, starting from concrete use cases in robotics and 3D asset creation instead of chat...

ACRouter treats routing as a self-evolving agent using a Context-Action-Feedback loop to learn from execution outcomes and dynamically select...

自托管SearXNG作为元搜索引擎，能聚合Google、DuckDuckGo等多源结果并保护隐私，让本地LLM免费获取网络信息，无需支付API费用。 仅需单条命令即可在Proxmox容器中快速部署，配合非推理模型即可实现自动化工作流中的实时数据校正。

世界模型正从LLM的数字局限中突围，成为通往物理AGI的核心路径。

• 上半年风险投资已超30亿美元，达沃斯论坛与广东规划均重点提及
• 与VLA不同，世界模型将视觉、认知、决策一体化训练，而非简单拼接
• 最大瓶颈在于交互式多传感器数据，需“边工作边采集”模式突破
• 2035年相关产业规模预计达10万亿美元，具身智能市场也将破万亿

SingGuard-NSFA tackles the shift from passive AI to autonomous agents by defending against prompt injection, goal hijacking, and privilege abuse. The...

腾讯通过姚顺雨重组混元研发体系，Hy3以MoE架构和工程化落地实现低调逆袭，上线三天即登OpenRouter第八并推高股价4.82%。这标志着腾讯从“大而全”同质化转向产品反馈闭环的造模型方法，但仍需连续迭代与微信原生Agent才能真正转化生态优势。

Model Releases & Technical Advances

• 🔥 GPT-5.6 Implications: Discussion on GPT-5.6 balancing efficiency and cost for open-source developers.
  -...

China showcases AI for global good at summits with family robots and video models, yet considers restricting top models like DeepSeek from overseas...

GPT-5.6 对开源开发者意味着一场效率与成本的重新平衡，推动闭源与开源生态调整。

多校position paper指出，平台受数据壁垒限制（竞争、监管、隐私、认知），永远只能看到用户局部行为；用户才是跨平台与线下完整上下文的唯一整合者。 LLM Agent能读取多源数据、推理偏好并执行推荐，使“用户主导个性化”从理论变为可操作实验结果显示跨平台数据显著提升排序与探索精度。 未来平台提供候选，用户Agent负责最终决策，二者互补而非取代。

• WAIC highlights embodied AI track with 200+ firms, shifting focus from models to real-world scenarios as industry AI penetration exceeds 80%.
  -...

George Hotz cheers real LLM gains like local GLM-5.2 handling tmux setup while blasting negative hype and takeover claims as manipulative fiction.

-...