Meituan releases LongCat-2.5: 1.6T-parameter multimodal MoE with 48B active per token

Meituan's LongCat team has released LongCat-2.5, a native multimodal mixture-of-experts (MoE) model with 1.6 trillion total parameters and approximately 48 billion activated per token. The model page is on Hugging Face under meituan-longcat. It builds on LongCat-2.0 and targets long-horizon agentic tasks: complex codebase development, full-project builds, creative design and 3D tasks.

Architecture and training

LongCat-2.5 builds on LongCat-Next for visual modeling. Native-resolution visual features are converted into discrete tokens and jointly aligned with text in a shared semantic space.

According to Meituan, training combines several elements:

• Agentic RL: real-world engineering tasks and reproducible agent failures, with efficient rollout strategies and safeguards against reward hacking.
• Multi-teacher learning: specialized agent, reasoning and interaction teachers guide the model through full-vocabulary KL distillation and Tree Attention.
• Long-context training efficiency: supernode-aware parallelism, efficient sharding and communication-computation overlap. Meituan reports 86% higher per-device throughput than the DDP baseline in reported long-sequence SFT tasks.
• RL infrastructure: aligned training and inference logic keeps token log probabilities bitwise-identical across training, prefill and decoding. KV cache optimization and LSA nearly double decode KV capacity without increasing TPOT on the tested workload.

Benchmarks

Scores marked with an asterisk are cited from public technical reports, blog posts or official leaderboards. Unmarked scores are measured in-house and are described by Meituan as indicative only. Comparison models include GPT-5.5, GPT-5.6-Sol, Claude-Opus-4.8, Claude-Opus-5, GLM-5.3, kimi-k3 and DeepSeek-V4.1-Flash.

Benchmark	LongCat-2.5	LongCat-2.0	Selected comparison points
Terminal Bench v4.0	38.5	3.0	GPT-5.6-Sol 37.3*, Claude-Opus-4.8 23.6*, Claude-Opus-5 53.9*
Program Bench (Almost Resolved)	25.5	2.5	GPT-5.6-Sol 23.0*, Claude-Opus-4.8 16.5*, Claude-Opus-5 37.0*
VideoMME	85.2	-	GPT-5.6-Sol 89.5*, Claude-Opus-4.8 86.0*, kimi-k3 90.0*

Terminal Bench v4.0 and Program Bench were run with the Claude Code harness (v2.1.280) at max_seq_length of 512K. VideoMME used at most 2,048 frames per video (1 fps) and at most 256 tokens per frame, with thinking mode enabled. Unless otherwise specified, LongCat-2.5 results use enable_thinking=on, temperature 1.0 and top_p 0.95.

W4A4 quantization

LongCat-2.5 applies W4A4 quantization to expert weights and activations, enabling native INT4 computation on domestic AI accelerators. It uses FlatQuant-style learnable transformations and, according to Meituan, maintains accuracy comparable to the unquantized model on evaluated benchmarks.

A shared expert-input transformation lets quantization happen before token replication and dispatch, so the same INT4 representation is reused across communication and computation. Against a W8A8 MoE baseline, online stress tests show:

• 43.3% higher prefill throughput
• 42.77% higher decode throughput
• 15.91% lower TTFT
• 19.72% lower TPOT

Grouped GEMM kernels achieve an average 1.73x speedup over W8A8. Combined with KV cache quantization, W4A4KV8 delivers 92% higher decode throughput than W8A8KV16.

Deployment and usage

LongCat-2.5 can be deployed on NPU platforms, with deployment instructions pointing to SGLang-FluentLLM. It is a hybrid model with a thinking mode and a non-thinking mode. Thinking mode (enable_thinking=True) with all reasoning content retained (save_history_reasoning_content=True) is recommended for better performance. Non-thinking mode (enable_thinking=False) is recommended for token efficiency. Context window and pricing are not yet disclosed.

What this means

LongCat-2.5 pairs a 1.6 trillion parameter MoE, with about 48 billion active per token, with an INT4 serving path built for domestic AI accelerators. The reported W4A4 gains are measured against a W8A8 MoE baseline, and the in-house benchmark scores are labeled indicative only by Meituan. On the benchmarks listed above, the largest gaps between LongCat-2.5 and LongCat-2.0 appear on agentic coding tasks, which fits the model's stated focus. Teams evaluating it should note that several comparison scores are cited from public reports and were not run under the same harness settings.