THE SCOREBOARD FOR WHAT'S NEXTAI + QUANTUM, SCORED. COMPARED. EXPLAINED.
AIQUANTUMSCORE.
← AI & Quantum News

AI / NEWS BRIEF

Meituan releases LongCat-2.5 for multimodal agent work

LongCat-2.5 moves from preview into Meituan's production API with visual understanding, extended coding workflows and compatibility with several agent harnesses.

By AIQuantumScore · · Event 2026-11-02

The production model replaces a September preview

Meituan's LongCat API changelog lists LongCat-2.5 as available on November 2, following the LongCat-2.5 Preview posted on September 25. The company positions the new release as a native multimodal model for visual coding, creative design, 3D tasks and long-running engineering assignments. Its API documentation lists compatibility with Claude Code, Codex, Hermes and other agent harnesses.

A large agent model with wider availability

TechNode reports that Meituan describes the model as a mixture-of-experts system with 1.6 trillion total parameters and approximately 48 billion activated per token. The report says it is available through LongCat's chat service and API, while GitHub and Hugging Face releases are planned to follow. Those architecture details and capability comparisons remain provider-supplied until independently reproduced.

Availability is not independent validation

OUR ANALYSIS

LongCat-2.5 is notable because it combines visual input and extended software work in a model now offered for active use rather than preview-only testing. Meituan says the model can sustain multi-step reasoning, tool use and full-project execution, but the public materials read for this article do not independently establish reliability, cost efficiency or benchmark leadership. AIQuantumScore has not tested the exact model, so no benchmark or REVAL score is assigned.

More to follow