Meituan’s LongCat team introduced LongCat-2.5 on November 1, a native multimodal model built for long-horizon agentic tasks. With 1.6 trillion total parameters and about 48 billion activated per token, it builds on LongCat-2.0 with stronger visual understanding and agentic coding, and it is live through the team’s chat service and programming interface, while its open-weight release on GitHub and Hugging Face remains marked as coming soon.
Seeing and coding in one model
The headline idea is native multimodality: LongCat-2.5 turns images into discrete tokens at several levels, so its backbone reads visual and textual inputs in one unified format. That design, the team said, removes the separate vision encoder from training and lets visual coding, creative design and three-dimensional tasks run through the same model that writes and reasons about code.
Eight demonstrations on the launch post show the ambition: real-time fluid simulation, three-dimensional exhibitions and city dioramas, business analysis and event operations, all generated end to end by the model. Demos are vendor-chosen, so they show the model at its best, not on an average task, but the range across software and office work is the point.
The training bet
Underneath the demos sits a training bet on long-horizon work. The team said it fed more long-horizon trajectories into long-sequence training and used large-scale reinforcement learning to strengthen engineering across codebases and full-project builds, converting real agent failures into verifiable training tasks. The goal is an agent that survives a long job, not one that shines on a short prompt.
Infrastructure is part of the pitch too. LongCat-2.5 was optimized for long-context training and four-bit inference on domestic Chinese accelerators, with the team reporting throughput and latency gains over an eight-bit baseline in its stress tests. That emphasis matters because it signals a model designed to scale without imported accelerators, even if the speed figures come from the vendor’s tests.
Scores, with asterisks
The launch post’s benchmark table puts LongCat-2.5 at 88.4 on Terminal Bench version 2.1, up from 70.8 for LongCat-2.0, and at 91.9 on BrowseComp, with visual scores of 55.4 on a text-to-3D test and 78.6 on Vision2Web, as reported by the team.
Rival scores marked with an asterisk come from reports or leaderboards, the team said, while unmarked rival scores are its own runs and indicative only. That caveat matters because benchmark tables invite horse-race reading, and this one is a vendor claim, not an independent verdict — the caution behind the benchmark where the best model scored just 12 percent.
Our read
LongCat-2.5 lands in a crowded Chinese agent race, next to DeepSeek’s autonomous coding push and Moonshot’s Kimi K3.1 Pro, whose full weights are due by November 4. Meituan’s pitch differs in emphasis: less about raw chat ability, more about an agent that can carry a long engineering job from vision to working software.
The open-weights question is the real test. LongCat-2.0 shipped with open weights, but 2.5’s buttons still read coming soon, so outside researchers cannot yet verify the training or benchmark claims for themselves. If the weights arrive, independent testers will quickly learn whether the visual-coding claims survive outside Meituan’s demos.
Why It Matters
If LongCat-2.5 performs as claimed, it narrows the gap between Chinese labs and the frontier on the hardest agent work: long, messy, multi-step engineering rather than short answers. It also shows how far domestic-hardware training has come, since the model was built for accelerators made in China.
For developers, the practical question is narrower: whether an agent that sees designs and code together can finish real projects with less hand-holding, and whether its long-context training holds up across a full codebase rather than a single file today.
Background
LongCat is Meituan’s AI model family, and LongCat-2.0 arrived earlier in 2026 as an open-weight model aimed at agentic coding. Version 2.5 keeps the same 1.6-trillion-parameter scale and roughly 48 billion activated parameters per token, but adds native visual understanding that 2.0 lacked, moving the family from text-only agents toward agents that can read a design and build it.
Source: LongCat launch post and full launch report
0 Comments