Tech
Meituan opens LongCat-2.5 1.6T MoE agent model under MIT, self-hosting guide names 194 Huawei Ascend accelerators and no Nvidia GPU.
The Decoder (Jonathan Kemper) · 2026-11-15 · Source ↗
Meituan opens LongCat-2.5 1.6T MoE agent model under MIT, self-hosting guide names 194 Huawei Ascend accelerators and no Nvidia GPU.
The Decoder (Jonathan Kemper) · 2026-11-15 · Source ↗
Jonathan Kemper's Sunday write-up in The Decoder lands one week after Meituan unveiled LongCat-2.5 and now puts the weights, the quantized variant, and the self-hosting path in public hands. LongCat-2.5 is a 1.6-trillion-parameter mixture-of-experts model, with about 48 billion active parameters per token, a 1-million-token input context, a 128,000-token output limit over the API, and reasoning that can be toggled on and off. It ingests text, images, and video and emits text; images are converted into tokens at native resolution rather than through a separate encoder. The model inherits from LongCat-2.0 and from the earlier research model LongCat-Next, and is explicitly positioned for long, multi-step agent work, coding above all.
Trained on Chinese hardware, deployed on Huawei Ascend
Meituan describes LongCat-2.5 as trained "entirely on Chinese hardware" without naming the training cluster. Kemper flags LongCat-2.0's prior disclosure — more than 50,000 Chinese AI chips in late June — as the closest public reference. The reinforcement-learning pipeline runs tasks built from real GitHub pull requests and recreations of past agent failures; training environments block internet access and strip Git history and cached files so the model can't cheat by remembering the answer, and an LLM checks solutions for shortcuts. Capabilities were distilled from several specialized teacher models covering agent work, reasoning, and dialog.
The deployment side is where the Nvidia story sits. Meituan's self-hosting guide, built on top of the SGLang-FluentLLM NPU runtime, names 194 Huawei Ascend Atlas A2 accelerators in its sample configuration: two accelerators for image processing, 64 for prompt ingestion, 128 for response generation. The W4A4 quantization — 4-bit weights, 4-bit activations — is native on Chinese chips; against the 8-bit baseline, Meituan's own load tests show prompt processing up 43.3 percent and response generation up 42.77 percent while accuracy stays "on par with the uncompressed model."
"On par with the uncompressed model." — Meituan, on the W4A4 quantization in LongCat-2.5, per The Decoder, 2026-11-15
Benchmarks and the DeepSWE jump
The Terminal Bench v2.1 score of 88.4 sits beside Claude Opus 5 at 89.1, GPT-5.6 Sol at 88.8, and DeepSeek-V4.1-Flash at the table top of 90.6. On DeepSWE LongCat-2.5 jumped from 9.7 on LongCat-2.0 to 64.6, behind DeepSeek-V4.1-Flash at 74.2 and GPT-5.6 Sol at 72.7 but ahead of Claude Opus 4.8 in the published table. BrowseComp topped the chart at 91.9 but was run in Meituan's own environment. Claude Opus 5.5 and GPT-6.1 Sol are not in the DeepSWE comparison, a conspicuous absence.
Distribution
Weights are MIT-licensed, allowing commercial use, modification, and redistribution. A separate W4A4 variant is also published. The chat endpoint lives at longcat.ai; the API supports both OpenAI and Anthropic request formats and works with agent harnesses including Claude Code, Codex, and Hermes. Launch promotional pricing is about €0.27 per million input tokens and €1.07 per million output tokens — a steep discount against Western frontier models. LongCat-2.0 ran anonymously as "Owl Alpha" on OpenRouter before its public release, which Kemper notes as a pattern for how Chinese labs are now timing the open-weight drop.
"The self-hosting guide has no Nvidia GPU guidance anywhere." — The Decoder, paraphrasing Meituan's documentation, 2026-11-15