AI 精选动态
智能评分 60
Ling & Ring 2.6 技术报告发布，开源基础模型
AI 推荐理由
相比前代，新增混合线性注意力与KPop稳定RL，在SWE-bench和ARC-AGI-2上取得显著成绩，值得阅读技术报告了解具体方法。核心解读
Ling & Ring 团队发布2.6版技术报告，开源两个基础模型。模型采用7:1混合线性注意力架构和KPop稳定代理RL，在SWE-bench Verified上达76.28%，token效率提升约4倍。具体型号：Ling-2.6-flash（104B，2.2x预填充吞吐量 vs Nemotron-3-Super）、Ling-2.6-1T（万亿级，AAII 34）、Ring-2.6-1T（深度推理，ARC-AGI-2 66.18）。
全文
Ling & Ring 2.6 technical report is out, with two open-weight base models.
We co-design model + system across architecture, training, and agentic capability:
• 7:1 hybrid linear attention
• KPop for stable agentic RL: SWE-bench Verified 76.28%
• ~4× token efficiency https://t.co/vMuUEOYXi9
Ant Ling (@AntLingAGI): 📑 Full technical report: https://t.co/pYvh8xKE8j
🧱 Ling-2.6-1T-base:
https://t.co/UedVW0lYFY
🧱 Ring-2.6-flash-base:
https://t.co/S2LG785WAU
💻 Code:
https://t.co/9oltGLtBpw
⚙️ Inference:
https://t.co/cc5Bzi1mH8
Ant Ling (@AntLingAGI): Three models. One architecture. Two open base checkpoints.
→ Ling-2.6-flash · 104B, instant execution, 2.2x prefill throughput vs Nemotron-3-Super
→ Ling-2.6-1T · trillion-scale instant flagship, AAII 34
→ Ring-2.6-1T · deep reasoning + long-horizon agentic, ARC-AGI-2 66.18
We made Ling-2.6-flash-base and Ling-2.6-1T-base available to facilitate more research oriented use cases! Feedback and use cases welcome!