Meituan open-sourced the weights of LongCat-2.5 on November 15, 2026. The model has 1.6 trillion parameters, uses a MoE architecture, and was trained entirely on Huawei Ascend chips without using any Nvidia GPUs.
Fact Recap
According to a report by the-decoder.com, LongCat-2.5 is a long-horizon task model built for AI agents, with roughly 48 billion activated parameters, supporting a context window of up to 1 million tokens and generation of up to 128,000 tokens, and accepting text, image, and video inputs with text output. The weights are released under the MIT license, and the API is compatible with OpenAI and Anthropic formats.
It scores 88.4 on Terminal Bench v2.1, compared with 89.1 for Claude Opus 5 and 88.8 for GPT-5.6 Sol; on the DeepSWE benchmark it improved from 9.7 in the previous generation to 64.6, still trailing Deepseek-V4.1-Flash at 74.2 and GPT-5.6 Sol at 72.7. It scores 38.5 on Terminal Bench v4.0, above GPT-5.6 Sol's 37.3.
Mechanism Breakdown
LongCat-2.5 converts images at native resolution into discrete tokens and processes them together with text, avoiding an additional image encoder. In the reinforcement learning stage it uses real GitHub pull request tasks, and in the sandbox it removes network access, Git history, and cached files, with a language model validating solutions to prevent shortcuts. Capabilities from multiple teacher models were distilled during training.
At inference, expert weights and activations are quantized to 4bit, raising prompt-processing throughput by 43.3% and response generation by 42.77% versus the 8bit version, with accuracy held flat. A self-hosted deployment requires 194 Huawei Atlas A2 series Ascend accelerators, of which 2 handle images, 64 handle prompts, and 128 generate responses.
Industry Impact
Previously, Meituan's LongCat-2.0 had already completed training entirely on more than 50,000 domestic AI chips by the end of June 2026. This open-source release marks the first time a Chinese company has publicly disclosed a complete domestic-chip training path at flagship scale under export controls. The MIT license permits commercial use, modification, and redistribution, and the API is offered at promotional prices of about 0.27 euros per million input tokens and 1.07 euros for output.
Strategic Assessment (Analysis, Not Fact)
From an evaluation standpoint, the domestic-chip training path has already approached or surpassed contemporary Western models on some agent benchmarks, though an overall gap with top models remains. This may lower the barrier to entry for Ascend hardware in overseas markets while offering Chinese companies a viable path to sustain frontier model iteration in a constrained environment. Future independent third-party cross-comparisons will be needed to confirm the capability ceiling.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接