Meituan launches LongCat-2.5 with 1.6T parameters and discrete visual tokens

Meituan published the LongCat-2.5 technical report on 1 November 2026 and enabled API access on 2 November 2026. The Mixture-of-Experts architecture totals 1.6 trillion parameters with roughly 48 billion active per token. Instead of bolting a continuous vision encoder onto a text backbone, Meituan pre-tokenises native-resolution images offline through Residual Vector Quantisation into discrete visual tokens, removing the online vision encoder from training and balancing multimodal batches across sub-80 GB domestic AI accelerators.

Post-training combines asynchronous partial-rollout reinforcement learning with Multi-teacher On-Policy Distillation using Tree Attention across Claude Code, Codex and Hermes agent harnesses. For serving, native INT4 W4A4 quantisation raises prefill throughput by 43.3% and decode throughput by 42.77% over W8A8 while cutting time per output token by 19.72%. Meituan draws its score tables interactively on the report page rather than static chart files. LongCat-2.5 is live on longcat.ai/chat and the OpenAI- and Anthropic-compatible api.longcat.chat endpoint; open-weight releases on GitHub and Hugging Face are listed as coming soon, and per-token API rates are not printed in the report.

Key numbers

• 1.6 trillion total parameters with roughly 48 billion activated per token.
• 88.4 on Terminal Bench v2.1, 91.9 on BrowseComp and 83.1 on MCP Atlas.
• 55.4 on 3DCodeBench Text-to-3D visual-programmatic synthesis.
• 43.3% higher prefill throughput and 42.77% higher decode throughput with native INT4 W4A4 versus W8A8.