Meituan LongCat-2.5 Explained: 1.6T Parameters, Benchmarks

Meituan, the Chinese delivery giant, released LongCat-2.5 on November 2, 2026. It is a native multimodal mixture-of-experts model with 1.6 trillion total parameters, about 48 billion active per token, built for long-horizon agent work such as coding across whole codebases.

Quick summary

• Size: 1.6T total parameters, roughly 48B active per token.
• Context: 1 million tokens in, up to 128,000 out (per TechNode-linked reports).
• Access: Meituan's chat service and API now; GitHub and Hugging Face weights are promised later.
• Benchmarks: all vendor-reported, not independently verified yet.

What is LongCat-2.5?

It builds on LongCat-2.0 and adds native vision. Instead of bolting on a separate image encoder, it turns visual input into multi-level discrete tokens so one backbone handles text and images. Meituan says this supports visual coding, design and 3D tasks.

How was it trained for agent tasks?

Meituan describes an agentic reinforcement-learning pipeline that auto-generates training tasks and test cases from GitHub and internal scenarios, and turns real agent failures into reproducible tasks. To limit reward hacking it cuts network access, clears Git history and uses an LLM judge. It also says the model was trained on domestically produced Chinese accelerators with under 80 GB of memory per card.

What are the benchmark numbers?

Meituan says its Terminal Bench v4.0 score beats GPT-5.6-Sol's 37.3. GeekPark's overnight roundup notes the company claims wins over Claude Opus 5 and GPT-5.6-Sol on its own agent tests while still trailing Opus 5 on Terminal Bench v4.0. Treat these as marketing claims until third parties reproduce them.

How is inference made cheaper?

Meituan reports a W4A4 scheme that quantises expert weights and activations to 4-bit, lifting prefill and decode throughput by about 43% against an INT8 baseline, with lower time-to-first-token.

Why does it matter?

It is another sign that Chinese labs compete on cost and agent performance, alongside DeepSeek's coding agent and rivals like Gemini 4 Carbon. For Indian developers, the real test is whether the open weights arrive and what API pricing looks like; neither is confirmed here.

Related articles

Mistral Large 4 Open Weights: Benchmarks and What It NeedsGDPx AI Benchmark: Claude Opus 5.5 Tops at Only 12%GPT-6.1 Terra: Price, Context and How It Compares to SolAnthropic IPO Explained: ANT Ticker, Revenue and Date