JOURNAL / 2026.11.03

LongCat-2.5 combines vision and development agents: what changes after the preview

Meituan releases LongCat-2.5 with training for long tasks and API access. Its interest lies in connecting images, code and verification; the preview already understood images, and reliability across the full cycle still needs testing.

A screenshot can describe a problem that a bug report misses: a menu covers the save button, even though the form works. An agent that receives that image, finds the responsible component and then verifies the repair could complete an important part of development work. That is the use I find interesting in Meituan's LongCat-2.5 release. It is a hypothetical example, not a test we have run with the model.

The official changelog dates LongCat-2.5 availability to November 2. It also preserves a crucial antecedent: LongCat-2.5-Preview already offered image understanding on September 25. This week's change should therefore not be described as the first time this version can see. It marks a new stage of the model and a fuller account of how it is trained to sustain engineering tasks across multiple steps.

The technical announcement describes representing images as discrete tokens that the model processes alongside text. For agents, it explains reinforcement learning with environments, tools and verifiable tests, as well as turning real failures into new training tasks. The valuable idea is to use the points where an agent loses track, misinterprets a result or declares the work complete too soon.

The Meituan team's publication positions that improvement over LongCat-2.0 for complex repositories, full projects and visual programming. These are the developer's stated goals. Including more long trajectories in training is a plausible explanation for progress, but does not establish how much an unrelated project will improve. Nor does the label "native multimodal" guarantee that the system will notice the visual detail that matters to a person.

Access does have verifiable consequences. Vercel announced its addition to AI Gateway on November 2, with text and image inputs. This provides another documented integration route. It confirms a service offering, without constituting an independent capability evaluation. LongCat's direct documentation also accepts images and videos; videos are submitted by URL. Teams should check modalities and limits for the particular route they intend to use.

Two details affect that integration. The parameter reference defines the roughly million-token context as the sum of input and output. It is not a million reserved for documents with all generation added afterwards. It also fixes temperature according to whether reasoning is enabled or disabled and overrides other submitted values. Apparently identical settings across providers can end up producing different experiments.

Return to the menu covering the button. I would ask for three separate results: locate the defect in the screenshot, correct the code and demonstrate that the interaction works again at the affected screen size. An attractive final image does not show that saving works; a functional test does not show that the button is visible. The agent's value would lie in preserving both conditions throughout the repair.

I would add one variant with an outdated screenshot and another with insufficient information to reproduce the error. I want to know whether it requests context or modifies an innocent component to satisfy an incorrect description. In long tasks, continuing without a good reason can cost more than stopping to ask. The criterion would be how many repairs a reviewer accepts, with what regressions and how much additional work.

My reading favors this direction: bringing perception and execution closer could expand the work we delegate to an agent. The demands also increase. Each step adds an opportunity to confuse a convincing representation with a correct outcome. LongCat-2.5 deserves a trial where image, code and behavior form a single task with explicit acceptance conditions. The release and its documentation let us design that trial; we still need its result.

Primary sources consulted on November 3, 2026. We have not run LongCat-2.5 or reproduced its evaluations. The interface example and its variants are our own proposals.

← Back to journal