Meituan LongCat-2.5-Preview: 1.6T Parameters, API Only — How to Use It Without Getting Locked In
📖 Glossary
AI Box (also known as Agent Computer / Agent PC), is a dedicated local hardware device that runs AI Agents. Pre-installed with an AI agent management system, plug-and-play, running 24/7. Users can remotely command AI to work via Discord, Slack, Telegram, WhatsApp, and more.
Meituan LongCat-2.5-Preview: 1.6T Parameters, API Only — How to Use It Without Getting Locked In
KAIHE has been saying one thing from day one: models are tools, the base is yours. On September 25, Meituan's LongCat API platform launched LongCat-2.5-Preview — 1.6 trillion parameters, 1M token context, native multimodal understanding, built for long-range agent tasks. But there's no open-weight release this time. Three months earlier, the same company put LongCat-2.0's weights on HuggingFace under MIT license. This generation? Nothing.
This isn't Meituan changing its mind. It's the industry direction shifting. When the API leg can make money, the weight leg doesn't need to walk anymore. For developers, binding to a single API means you accept price changes, you're passive when service stops, and your data flows through someone else's servers. KAIHE AIBOX agent computer does local orchestration: cloud models are called as tools, workflows stay on your device, and swapping models is like swapping batteries.
Same Scale, New Eyes and Hands
LongCat-2.5-Preview continues the MoE architecture with approximately 1.6 trillion total parameters and roughly 48 billion activated per inference — under 3% activation rate, deliberately trading density for inference cost efficiency on long-range tasks. Native 1M token context with 128K max output means it can ingest an entire codebase or a full financial report in one pass.
Two changes stand out. First, native image understanding — the model can parse image content, do cross-modal Q&A, content summarization, and complex visual reasoning. Second, explicit positioning for "long-horizon" tasks: Meituan lists terminals, browsers, desktop software, spreadsheets, and design tools as target scenarios. The model operates across multiple interfaces within a single task, not just answering one question.
Coding capability is the marketing focus. Meituan claims strong performance in code generation, code understanding, and automated programming, with compatibility for Claude Code, OpenCode, OpenClaw, Hermes, and Kilo Code. The API supports both OpenAI and Anthropic formats — change one base URL and key, and you've switched backend models.
Previous Generation Open-Source, This Generation API-Only
Timeline: March saw LongCat-Next for long-context validation with "open-source model weights and inference code." June 30 brought LongCat-2.0 — fully open-sourced under MIT license on GitHub and HuggingFace, hailed as China's first trillion-parameter model trained end-to-end on roughly 50,000 domestic compute cards. September 10: showcased at CIFTIS. September 25: 2.5-Preview launches — fifteen days later.
But 2.5 ships API-only. No weights, no open-source mention in the changelog, no benchmark scores. The previous generation's model page proudly linked to GitHub and HuggingFace repositories. This one doesn't even have a link.
The logic is straightforward: when the API leg generates revenue, giving away weights becomes optional. Pricing: limited-time rates at $0.30 per million input tokens, $1.20 output, $0.006 cached input. Chinese domestic rates: ¥2 uncached input, ¥0.04 cached, ¥8 output. Existing users got 5 million free tokens. Prices match the previous generation — but "matching" is a Preview-stage promise; official pricing awaits formal launch.
The Long-Range Agent Track Is Crowded
The same week, Xiaomi disclosed MiMo-V3's core HySparse2 architecture, also optimized for long-horizon multi-turn agent inference. Google's Gemini 3.8 Live just launched with real-time voice and video reasoning. OpenAI is pushing the "o" always-on assistant while pausing frontier training. Meituan, Xiaomi — companies labeled "second-tier" — are all betting on the agent long-range task track.
LongCat-2.5's positioning shifted from "general capability追赶" to "autonomous software interaction" — executing multi-step tasks across terminals, browsers, and GUIs. This complements what KAIHE AIBOX does: KAIHE handles orchestration, LongCat serves as one called model, and workflows remain on local devices.
The Risk of API Lock-In
Meituan hasn't published benchmarks. How much LongCat-2.5 improves over 2.0 on GUI operations and long-horizon tasks is unknown. More importantly, binding your long-range agent workflow to LongCat's API means three things: First, your data transits Meituan's servers — codebases, logs, documents all pass through their cloud. Second, you accept price changes; on May 29, Meituan unilaterally retired six older model APIs. Third, Preview has no SLA guarantee — if it goes down in production, you're on your own.
KAIHE's approach brings control back locally. OpenClaw handles task orchestration, Hermes executes actions, and cloud models — whether LongCat, GPT, Claude, or DeepSeek — are called as tools. Swap a model API, and the workflow doesn't need rewriting. Prices go up, auto-switch to a cheaper one. Service stops, swap another. Your data, your workflow, your scheduling logic all stay on your local device. Unplug and it stops. Physical isolation.
This isn't opposing LongCat — it's cheap, has long context, and Anthropic-format compatibility keeps integration cost low. As one called model among several, it's a good option. But "a useful tool" and "binding your entire workflow to it" are two different things.
Further Reading
- OpenAI Paused Training Twice While Racing to Launch "o" — The less controllable the cloud, the more critical the local base
- DeepSeek API Price Hike: The Era of Binding to One Model Is Over — Price hikes aren't the scary part — lacking switching capability is
- Kaihe AIBOX A1: What Is a Local Agent Computer — Hardware specs and local orchestration architecture
Learn more — search [KAIHE AIBOX]
Contact: [email protected]
KAIHE AIBOX · Your 24/7 private AI assistant