longcat-2.5: Meituan Bets on Long-Horizon Agents

longcat-2.5 is Meituan's latest shot at the part of the model race that matters most to teams building agents: long-horizon tasks, where a model has to plan, use tools and keep working through a complex codebase instead of answering one prompt. Released on November 1, 2026 with a detailed technical blog post, it is a 1.6-trillion-parameter mixture-of-experts model with about 48 billion parameters activated per token, natively multimodal, and trained explicitly for agentic coding across multiple agent harnesses. This post breaks down what Meituan claims, how the training and inference story works, what the vendor-reported benchmarks do and do not tell you, and how it lands next to this week's other agent releases. Source

What LongCat-2.5 is

According to the announcement, LongCat-2.5 builds on LongCat-2.0 with stronger native visual understanding and agentic coding. Meituan describes folding more long-horizon trajectories into long-sequence training and using large-scale long-horizon reinforcement learning to improve end-to-end engineering on complex codebases and full-project builds. Two details stand out for practitioners. First, the training pipeline converts real-world agent failure trajectories into reproducible, verifiable training tasks, so the model keeps getting targeted data for its actual weak spots rather than generic fine-tuning. Second, the model is trained and evaluated across harnesses: Meituan lists compatibility with Claude Code, Codex and Hermes on its homepage, which matters because a model that only performs inside one vendor scaffold is hard to adopt. We covered the harness question in depth in our post on harness2 and agent harness tuning. Source

Native multimodality through discrete visual tokens

Rather than bolting a vision encoder onto a text model, LongCat-2.5 represents visual signals as multi-level discrete tokens, so the backbone processes images and text in one unified token format. The design applies Residual Vector Quantization to native-resolution features from a semantic-aligned encoder, inherits the approach from Meituan's LongCat-Next work, and learns visual codebook embeddings jointly with the backbone. There is a practical benefit beyond elegance: because visual token IDs can be precomputed and cached offline, the training system avoids running a vision encoder in the training graph, which simplifies load balancing across long sequences. Meituan says the model uses this for visual coding, creative design and 3D tasks, and its showcases include a real-time fluid simulation webpage and 3D scenes generated end to end. Source

The agentic RL pipeline

The reinforcement learning section is the meatiest part of the post for AgentOps readers. Meituan describes a synthesis pipeline that filters GitHub pull requests and linked issues, uses agents to annotate tasks and build interactive environments, and validates them with fail-to-pass and pass-to-pass tests, with live rollout pass rates guiding relabeling. Training uses asynchronous partial rollouts: in-flight trajectories are interrupted, saved and resumed to keep batches full, and the team clips the discrepancy between sampling-policy and current-policy log-probs to bound gradient bias from stale tokens. The anti reward-hacking measures are worth quoting as a checklist for anyone building agent evals: network access is disabled in task environments, git history and cached files are removed, and an LLM judge audits trajectories for exploits, with extra penalties for repetitive reasoning and repeated tool calls. These are vendor-described mitigations, not guarantees, but they map closely to the failure modes we see when auditing agent behavior in production. Source

The benchmarks, with the right caveats

Meituan's own reported results show LongCat-2.5 at 88.4 on Terminal Bench v2.1, up from LongCat-2.0's 70.8, and 38.5 on the harder Terminal Bench v4.0, ahead of the 37.3 listed for GPT-5.6-Sol but behind the 53.9 listed for Claude-Opus-5. On general agent tasks it reports 91.9 on BrowseComp, 83.1 on MCP Atlas, 51.8 on AutomationBench and 65.4 on JobBench; on visual tasks, 78.6 on Vision2Web and 96.5 on MathVision with tools. Now the caveats, and they matter. These are vendor-reported numbers. In Meituan's table, many competitor scores carry an asterisk meaning they come from public technical reports, blog posts or official leaderboards, while others are self-run and, in Meituan's own words, indicative only. Cross-model comparisons here mix harnesses, judges and settings, and we did not find independent verification of these scores. The honest read: LongCat-2.5 looks highly competitive in long-horizon coding on its maker's own tests, and the leap over LongCat-2.0 is large and internally consistent, but treat any ranking against GPT or Claude as provisional until third-party evals exist. Source

Infrastructure: domestic accelerators and W4A4 inference

A 1.6-trillion-parameter model is expensive to serve, so the inference story is part of the product. Meituan reports a W4A4 scheme that quantizes both expert weights and activations to 4 bits and runs the matrix multiplications on native INT4 units, with accuracy comparable to the unquantized model on all of its evaluated downstream tasks. In Meituan's online stress tests, prefill throughput improves by 43.3 percent and decode throughput by 42.77 percent versus a W8A8 INT8 baseline, while TTFT and TPOT fall by 15.91 and 19.72 percent respectively, and the grouped-matrix kernel averages a 1.73 times speedup. Training ran on domestic AI accelerators with less than 80 GB of memory each, using supernode-aware parallelism and a dual-shard FSDP setup that Meituan says delivers 86 percent higher per-device throughput than a DDP baseline on long-sequence SFT. For buyers, the subtext is cost pressure on serving huge MoE models, and a signal that Chinese labs are optimizing around export-constrained hardware rather than waiting for it. Source

Availability

Per reporting from AIBase, the online demo and API access are open now, and GitHub and HuggingFace weights are slated for release soon; Meituan's own API documentation for the LongCat platform is already live. Until the weights land, evaluating the model means using Meituan's hosted endpoints, which adds the usual data-residency and vendor-lock questions for teams outside China. If open weights do arrive, the W4A4 work becomes much more interesting, because self-hosting a 1.6-trillion-parameter model at 4-bit precision is a different cost conversation than renting API tokens. Source

Where it fits, in our assessment

This section is our read, separate from what the sources say. LongCat-2.5 arrives days after DeepSeek's dsh-superagent prototype, and the pattern is hard to miss: the frontier labs in China are now competing on agent task effectiveness rather than chat benchmarks, and they are publishing the training plumbing to prove it. For engineering leaders, three things are worth doing. First, if you run a coding-agent stack, add LongCat-2.5 to your eval shortlist, especially for visual-to-web and full-project builds, but run your own tasks rather than trusting the vendor table. Second, watch the harness-compatibility claim in practice; a model that genuinely performs across Claude Code and Codex gives you negotiating room with both vendors. Third, do not migrate anything on day one: the weights are not out, the benchmarks are self-reported, and the model is one week old. The competitive pressure this puts on OpenAI and Anthropic pricing is real, but your roadmap should follow your evals, not a launch post. Source