LongCat-2.5 Adds Native Multimodality for Agentic Tasks
agents multimodal
| Source: HN | Original article
LongCat-2.5, a native multimodal model for long‑horizon agentic tasks, boasts 1.6 trillion parameters with roughly 48 billion activated per token.
Meituan’s AI research team has unveiled Long Cat‑2.5, a 1.6‑trillion‑parameter mixture‑of‑experts model designed from the ground up for multimodal, long‑horizon agentic tasks. The new system activates roughly 48 billion parameters per token and can process up to a million tokens of context, a step up from its Long Cat‑2.0 predecessor. Its chief upgrades are native visual understanding—handled through discrete visual tokens—and enhanced agentic coding capabilities, both trained with large‑scale reinforcement learning aimed at autonomous tool use and planning.
Benchmark scores released alongside the model show strong performance: 91.9 on BrowseComp, 88.4 on Terminal Bench v2.1 and 82.1 on PaperBench. Meituan also disclosed a pricing model of $1.2 per million output tokens, positioning Long Cat‑2.5 as a commercially viable option for developers building AI agents that must reason over extended sequences and interpret visual inputs.
The launch matters because it pushes the frontier of “agentic” AI—systems that can autonomously execute complex, multi‑step workflows. As we reported on the emergence of the agentic web and the need for deterministic safety gates in such workflows, a model that natively blends vision and code while maintaining a massive activation budget could accelerate the deployment of more capable autonomous agents. Its long‑context capacity also aligns with recent benchmarking efforts, such as SemiAnalysis’s GDPx suite, which evaluates agents across domains like chip design and cybersecurity.
Going forward, observers will watch how Long Cat‑2.5 performs on emerging agentic benchmarks, whether its cost structure spurs broader adoption, and how safety mechanisms are integrated to keep long‑horizon reasoning under control. The model’s reception could shape the next wave of AI‑driven automation across software development, visual programming and beyond.
Sources
Back to AIPULSEN