Prefrontal与苏黎世联邦理工开源LSA稀疏注意力技术
Prefrontal与苏黎世联邦理工学院研究者开源LSA，用语义分组加速超长上下文稀疏注意力。其在GLM和Qwen测试中报告显著预填充及首Token延迟改善，但多级索引准确率仍有下降。
Loading…
bibibi.ccbibibi.ccLoading…
Connect scattered updates into a story you can understand.
Prefrontal与苏黎世联邦理工学院研究者开源LSA，用语义分组加速超长上下文稀疏注意力。其在GLM和Qwen测试中报告显著预填充及首Token延迟改善，但多级索引准确率仍有下降。
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。
As of October 27，Token the expenditure price index stood at 0.94 USD/million Token, down from days earlier 1.9%, and down from its period high 54.48%。
DiffuSpace released and open-sourced a diffusion model based on a modification of Qwen3.6-35B-A3B Clover-35B-A3B-Preview. Its disclosed internal tests reportedly showed that the model reached GB300 on a single GPU 680 Token/s。
Anthropic is reportedly preparing a new Claude model that is better at reasoning and understanding user needs, but the related assessment comes from only early tester, with no public test data available. The community speculates that it may be Claude Fable 5.5。
NaiveAI releases N0.5-Flash and claims a peak single-stream inference speed of 2122 tok/s、AI participating in research and development. Moonshot AI researcher Yang Xinyu questioned its benchmark and architectural innovation claims, believing this was insufficient to prove that “AI creation AI” delivered a unique breakthrough.
Researchers from Stanford University and NVIDIA Research have open-sourced CLM-8B for rapidly selecting from candidate actions; official tests reportedly show it is up to approximately 9 times faster than Jev.