Meituan releases LongCat-2.5 weights, a 1.6-trillion-parameter agent model

For the head of model risk at a us bank

Meituan's LongCat team released the weights for LongCat-2.5, a 1.6-trillion-parameter model for coding and other long-running agent tasks, one week after its official unveiling, The Decoder reported on Nov. 15.1

Meituan's model card says the model builds on LongCat-2.0 and was trained with infrastructure optimizations on domestic AI accelerators; The Decoder reports it was trained entirely on Chinese hardware, and the sources do not say why Meituan released the weights.31

The model card describes a mixture-of-experts design with about 48 billion parameters activated per token, and the weights are under the MIT license, which allows commercial use, modification and redistribution, according to The Decoder.31

On Terminal Bench v2.1, Meituan's table lists LongCat-2.5 at 88.4 against 89.1 for Claude Opus 5; on the harder Terminal Bench v4.0 it lists 38.5 against 53.9.23

Meituan's model card says unmarked scores in its table, including LongCat-2.5's own, were measured in-house and are indicative only, while scores marked with an asterisk, including Opus 5's, are cited from public reports or leaderboards.3

The Decoder reports the API supports the OpenAI and Anthropic formats and works with Claude Code, Codex and Hermes, and that the self-hosting guide describes a cluster of 194 Huawei Ascend accelerators with no guidance for Nvidia GPUs.1

Why it matters

Free, MIT-licensed weights are simple to govern, but the self-hosting guide assumes a 194-chip Huawei cluster, so many teams may instead call Meituan's API, where the endpoint is a one-line config change. Banks and fintechs should set vendor, data-residency and approval rules for that route, and treat Meituan's self-run benchmark scores as unverified.

From the brief

Our take. A developer can point Claude Code at LongCat-2.5 for a promotional €0.27 per million input tokens. Your policy on Chinese-origin weights should exist before that change shows up in a config file.

Action. Model risk: Add Chinese-origin open weights and LongCat API endpoints to developer-tool allowlist rules and network egress blocks.

What our records say

Anthropic sells banks Claude seats and API access directly and through cloud marketplaces, and in 2026 added embedded engineering with Goldman Sachs and FIS.1 It also restricted cyber-defence access to its Mythos model for a few large banks and co-founded a services firm with Goldman Sachs.1 Customers: Goldman Sachs and JPMorganChase offer Claude to staff, and BankUnited and Flagstar run Claude in internal AI platforms.1

  1. Claude: AI programme summary

From our records

Sources

  1. ReportMeituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU The-decoder
  2. Primary sourcemeituan-longcat/LongCat-2.5-W4A4 · Hugging Face Huggingface
  3. Primary sourcemeituan-longcat/LongCat-2.5 · Hugging Face Huggingface