German AI lab Empero released Manta, claiming it can handle a 100000000 Token context; tests showed it performs well in long-document retrieval and programming record continuation.
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.
Bailing, the large model under Ant Group, announced the open-sourcing of Ling-3.1-flash, with approximately 560B total parameters and approximately Token activated per 25B, supporting up to 1M context; weights can be downloaded from Hugging Face and ModelScope.
Prime Intellect releases the open-source model inference service Prime Inference, with an API compatible with OpenAI API, and the first publicly deployed model is GLM-5.3. The company says the service processes nearly 1000000000000 tokens。
Germany AI company Aleph Alpha has released and open-sourced the English-German bilingual MoE model Kolibri; its weights have been uploaded to Hugging Face under the Apache 2.0 license. The material states that it supports tool calling, adjustable reasoning modes, and context of up to approximately 1000000 token s.
InclusionAI releases Ling-3.1-flash, with a clear increase in parameter scale and per-token activation scale compared with the previous generation. The official demonstration showed it completing long-duration programming tasks, with plans to make context and an open-source model available later.
OpenRouter The model page indicates that the anonymous model Space Bunny Alpha September 23 launched in preview, supporting million-Token context, multimodal input, and tool calling. Both input and output are currently free.