bibibi.ccLoading…
Connect scattered updates into a story you can understand.
Dongcha Beating AI reports that Meituan has released the LongCat-2.5 official version, focusing on improved visual programming and complex Agent task capabilities, and has opened online trial and API。
Doubao Work announced that it is opening third-party large language model services to Team Subscription and Enterprise Subscription customers, initially integrating the Kimi K3.1 Pro、GLM-5.3、DeepSeek V4.1 Flash and other models.
An independent developer built Qwen3.5-4B based on Qwen puter, integrating a 32-bit RISC-V virtual CPU and readable/writable memory into the model weights. The project can run compiled C programs, but its model size and computational efficiency remain clearly inferior to those of external tools.
Liquid AI Release LFM3 series of approximately 2000000000-parameter small models, and disclose that its 1.7B text model's webpage Agent test results exceed those of Qwen3.5-9B; it trails in complex programming and long-text tests.
Baidu released and open-sourced Unlimited-OCR 2, improving the efficiency of -page document parsing through sparse attention, chunked decoding, and cache reuse. Its results on Baidu's self-built benchmark exceeded those of the listed comparison models.
Mirendil releases Fern 10000000 Token with support for 0.1 context, and says its Agent wrote more than 99%of code during Mirendil-Base pretraining from scratch. The model is not yet available and can only be accessed by applying for early access.
BOSS Nanbeige Laboratory, under Direct Hire, has released and open-sourced Nanbeige4.3-33B, for programming and office work Agent, supporting 512K context and using the Apache 2.0 license. The team plans to distill its capabilities into subsequent 3B models.
A report by CAICT and other institutions shows that 2026 year June domestic Token average daily consumption was approximately 500000000000000, approximately 2025 year December that of 4 times.August The average of domestic models TPS rose to 78.23, but remained below that of typical overseas models.
Citing Moonshot AI's official blog, PANews states that Kimi K3.1 Pro and its high-speed version have launched across multiple products and API, with complete model weights planned to be made available before November 4. There is an inconsistency between the report date and the evidence publication date, pending further verification against the official original text.
London startup Artificial Societies has released Rothko to simulate the decision responses of consumers, industry professionals, and policymakers. The company says its survey prediction error is lower than that of 9 general-purpose large models.
PANews reported that Kimi（Moonshot AI）will release its latest large language model K3.1, and has released related teaser videos.
BlockBeats cited Max For AI as saying that Kimi K3.1 and Qwen4 may already be undergoing Arena anonymous testing and are expected to launch this week. Neither the identities of the models nor their specific release timing has been officially confirmed.