The announcement came via social channels, confirmed by the ModelScope team: the weights of longcat 2.0, a Large Language Model with 1.6 trillion parameters and roughly 48 billion active per token, have been released under the MIT license. A move that shifts the balance for anyone building on-premise inference stacks and aiming to avoid proprietary APIs or restrictive clauses.

This is no ordinary model. The stated architecture – a 1.6T giant with only 48B active – points straight to a Mixture of Experts (MoE) design, a scheme that keeps per-token compute costs in check by multiplying total parameters without blowing up FLOPs. In practice, for each token the model routes computation to only a fraction of the available experts, keeping latency comparable to that of a much smaller dense model.

From a deployment perspective, the flip side is memory. With 1.6T parameters to hold in VRAM, FP16 inference would demand over 3 TB of GPU memory – a figure that forces distributed solutions with multiple nodes and fast interconnects. The MIT license, however, removes all legal barriers to experimentation: one can aggressively quantize (INT8, INT4) without asking permission, and nothing prevents adapting the model to regulated enterprise environments, even air-gapped ones. For those with access to multi-GPU clusters – eight 80 GB A100s deliver 640 GB, a starting point after aggressive compression – longcat 2.0 becomes a concrete candidate for workloads ranging from code generation to document analysis in sectors where data cannot leave company boundaries.

The development team, as described in the technical blog published at the end of June, worked to balance capability and accessibility, and the choice of the MIT license – one of the most permissive – suggests a dissemination strategy aimed at the enterprise world without excluding the research community. In a landscape where many top-tier models remain locked behind SaaS interfaces or custom licenses, a release like this fuels the debate on which models can truly be managed in-house, at what cost, and with what guarantees of ongoing updates.