Flash

TopicLongCat-2.5 Release

Flash·Models & Products·2026-11-15 19:55

Meituan Open-Sources LongCat-2.5 Weights: 1.6T-Parameter Agent Model Trained Without Nvidia Chips

Meituan released the full LongCat-2.5 checkpoint on Hugging Face on Nov 15 under the MIT license, including for commercial use; the 1.6-trillion-parameter model activates about 48B per token with up to 1M tokens of context, one week after its official unveiling.

A sunlit AI data center hall with rows of white server racks, a large wall screen showing a colorful neural-network diagram, and a laptop displaying code on a white desk
AI-generated illustration

What happened

The model card for LongCat-2.5 now carries the full checkpoint on Hugging Face, letting developers download, inspect, modify, fine-tune, self-host and redistribute it under the MIT license. Built for long-horizon agent work, the model takes text, image and video input, outputs text, and lets users toggle reasoning on and off.

The release comes about a week after the model's official unveiling. The API supports both the OpenAI and Anthropic formats and plugs into agent tools such as Claude Code and Codex, with a chat interface at longcat.ai. A limited-time launch price is roughly €0.27 per million input tokens and €1.07 per million output tokens.

Key facts

  1. Scale: A native multimodal mixture-of-experts model with 1.6 trillion total parameters and about 48B activated per token; the API supports up to 1M tokens of context and generations up to 128,000 tokens
  2. Trained on domestic chips: Meituan says the predecessor LongCat-2.0 was trained in late June on a cluster of more than 50,000 domestic AI chips; the sample LongCat-2.5 deployment runs on 194 Huawei Ascend Atlas A2 accelerators — 2 for image processing, 64 for prompt ingestion and 128 for response generation — with no deployment guidance for Nvidia GPUs
  3. Terminal-task scores: In Meituan's own evaluation, Terminal Bench v2.1 scores 88.4, against 89.1 for Claude Opus 5, 88.8 for GPT-5.6 Sol and 90.6 for DeepSeek V4.1-Flash; on the harder Terminal Bench v4.0 it scores 38.5, above GPT-5.6 Sol's 37.3 and below Opus 5's 53.9; DeepSWE rises from 9.7 for the predecessor to 64.6
  4. Quantized deployment: Expert weights and activations use 4-bit quantization, which domestic chips compute natively; against the 8-bit version, online stress tests show 43.3% higher prompt-processing throughput and 42.77% higher decode throughput, with accuracy on par with the unquantized model

Why it matters

a 1.6-trillion-parameter agent model opens its full weights under the MIT license, trained and deployed entirely without Nvidia chips, with a Terminal Bench score of 88.4, close to Claude Opus 5's 89.1.
Useful Tap if this story helped you
Send to a friendMakes a share image for X, WeChat, Telegram and more

SourcesThe Decoder (Nov 15, 2026); Hugging Face model page (Nov 15, 2026). This article is compiled from public information and does not constitute investment advice.

From HIMIAO Horizon · himiao.ai

Grab todayOpenAI Codex for Students: $100 in Free CreditsHIMIAO Picks →

Follow HIMIAO Horizon

No sign-up · Unsubscribe any time

Email

Notifications

A ping on your phone or computer when the brief is out. No email needed.

HIMIAO PicksWeekly freebies: one email every Friday, a 5-minute read

Comments

  1. Loading comments…
Ask the cat