What happened
Meituan’s LongCat team describes LongCat-2.5 as a native multimodal model for long-horizon agentic tasks, building on LongCat-2.0. The release page is dated 2026-11-1 and characterizes the model as having 1.6 trillion total parameters and approximately 48 billion activated per token. The team says the model improves native visual understanding and agentic coding. The page offers a “Try it” link and an API access link; its GitHub and Hugging Face links are marked “Coming soon.” The supplied source excerpt does not state a general availability date, rate limits, or pricing.
The stated visual design represents visual signals as multi-level discrete tokens, so the backbone can process visual and textual inputs in a unified token-based format. Meituan says it combines native-resolution visual tokenization, offline preprocessing, and visual embeddings learned jointly with the backbone. Its description references residual vector quantization over native-resolution features from a Semantic-and-Aligned Encoder. It says fixed tokenization lets visual token IDs be precomputed and cached, simplifying sequence packing and distributed-training workload management. A preliminary embedding analysis, the company notes, was validated on a smaller MoE model with roughly 3 billion activated parameters on average; it is not a direct demonstration of LongCat-2.5’s production behavior.
For agentic training, the company says it scaled long-horizon reinforcement learning across data, environments, algorithms, and harnesses. Its task pipeline filters GitHub pull requests and linked issues, uses agents to annotate tasks and synthesize interactive environments, and validates tasks with fail-to-pass and pass-to-pass tests. The page also describes internal coding tasks, rollout pass rates that inform relabeling and filtering, and analysis of false-positive and false-negative trajectories to revise tests. For asynchronous partial rollouts, in-flight trajectories are saved and resumed; the team describes techniques intended to manage stale policy samples and preserve useful gradient signal.
The page includes a metrics dashboard comparing LongCat-2.5 with GPT-5.5, GPT-5.6-Sol, Claude-Opus-4.8, and Claude-Opus-5, and says broader results appear in a table. Its showcases are described as end-to-end outputs across software development and office scenarios, including a real-time fluid simulation, business analysis, 3D artifacts, and feedback reports. The page labels some demo footage as coming soon. These are company-reported comparisons and examples; the supplied source does not provide independent replication.
What the archive adds
Meituan's LongCat-2.5-Preview on OpenCode: 1M context, multimodal, offered free for a limited time covered a preview offer on OpenCode, not this release. This is a new event in the same product thread: the current announcement describes the full LongCat-2.5 model and its training and visual-processing approach.
Trend signal
This release supports scientific-agents-need-artifact-evidence only weakly: Meituan describes verifiable tests and failure trajectories as part of training, but this is a vendor account of model development, not an inspectable external evidence package demonstrating reproducible agent progress.
Why now
The announcement follows a preview that the archive says was offered free on OpenCode for a limited time. The new post presents LongCat-2.5 as the fuller model release and adds a technical account of its training pipeline and visual representation. The source does not explain why the release is timed now.
Connections to other events
The earlier LongCat-2.5-Preview offer on OpenCode connected the model to a coding environment and a limited-time access offer. This announcement is a separate development: it describes the model’s architecture and training, while the supplied material does not establish that the preview’s OpenCode terms carry over to the full release.
Why it matters
For teams evaluating agentic coding models, the useful part of this announcement is not just a parameter count. Meituan outlines how it constructs coding environments and tests, including fail-to-pass and pass-to-pass checks, and says real agent failures feed back into training tasks. That makes the verification design part of the model story: long-running coding work needs a way to tell whether an agent completed the requested change, not simply produced plausible code.
The release also presents a unified visual and text representation as a way to support visual coding alongside codebase work. Builders handling interfaces, diagrams, or visual assets may want to test that combination in their own workflow. But the provided page is a vendor description, and its dashboard and showcases are not independent evidence. Teams should treat the methods as evaluation hypotheses and test the model in their own harness, on representative repositories and visual tasks, with explicit completion checks. Later coverage: Meituan released LongCat-2.5, its new multimodal model with 1.6T total parameters and 49B active..
Limits and uncertainty
The supplied source is Meituan’s own release page. Its performance comparisons and showcases are company-reported; the excerpt does not provide independent evaluations, full benchmark methodology, confidence intervals, or enough detail to assess comparability across harnesses. It gives no API price, rate limits, availability geography, or operational service guarantees. GitHub and Hugging Face are marked as coming soon, while the page does link to a trial and API documentation. The source describes training techniques, but does not establish that they caused the claimed performance or that results generalize to a particular team’s codebase and tools.
Reader takeaway
LongCat-2.5 is a technically detailed model release: Meituan describes a multimodal token representation and a long-horizon agent-training pipeline that uses verifiable coding tasks. Treat the announced methods and scores as claims to test, not a substitute for a local evaluation. The practical next step is to check access and run representative tasks with explicit tests in the harness your team actually uses.