The full release is a training story

LongCat-2.5 existed before the full release. Meituan's API change log records the LongCat-2.5-Preview as available on September 25, with image understanding and coding capabilities. The November release adds a fuller technical description of its training, multimodal architecture and benchmark results.

The positioning is consistent across everything Meituan published with the release. TechNode's write-up leads with longer agent-action sequences and large-scale reinforcement learning aimed at more complex codebases and complete projects, rather than short, self-contained coding tasks. That is the same emphasis the September preview carried, when the change log described a model already working inside agent harnesses such as Claude Code and Codex.

Two training claims do the heavy lifting. The first is volume and length: agents trained on longer trajectories of actions, so the model learns to stay coherent across a task that takes many steps. The second, in Meituan's technical description, is that failures are converted into reproducible training tasks: the company describes mining GitHub pull requests and issues to build synthetic environments checked by fail-to-pass and pass-to-pass tests, and analyzing false-positive and false-negative trajectories to harden the training loop against reward hacking. That second claim is the more interesting one, because it describes a data engine rather than a bigger run: a system that mines its own errors for curriculum. It is also, for now, a claim. Meituan has published no ablation showing how much of the benchmark movement comes from failure-derived tasks as opposed to scale, data or harness choices.

The multimodal side is architectural rather than bolted on. Meituan says LongCat-2.5 tokenizes images into multiple levels of discrete visual tokens, quantized over semantic features, so pictures enter the same token stream as text and code instead of arriving through a separate vision encoder. The company connects that design directly to the tasks it is demoing: visual coding, 3D generation and screen-based work, where the model has to look at an interface and act on it in the same pass.

Read the scoreboard in two halves

Meituan's official table reports gains that point in the direction the training story predicts. It lists 88.4 on Terminal Bench v2.1, against 70.8 for LongCat-2.0; 65.1 on AutomationBench v1.0, against 51.8; 91.9 on BrowseComp; and 82.1 on PaperBench, against 45.2 for the older model. New multimodal rows appear too: 55.4 on 3DCodeBench (Text→3D) and 78.6 on Vision2Web, on tasks the text-only LongCat-2.0 could not attempt natively. These are Meituan's numbers, from Meituan's own table, and Pandaily notes the results have not been independently verified. Treat them as the company's account of its own progress, not as a settled ranking.

Then the table slows down. On Program Bench, which tests building complete programs from a specification rather than patching existing ones, Meituan's table gives LongCat-2.5 a score of 25.5, against 37.0 for the top entry in the same comparison, Claude Opus 5. That is not a rounding difference; it is the model trailing by roughly a third on exactly the kind of long, structured build task the release is marketed around. The pattern repeats on the table's other build-from-specification rows: 60.5 on NL2Repo against 75.3, and 38.5 on Terminal Bench v4.0 against 53.9. The gains half of the table is real, on the company's own evidence. So is the gap half. A release that only quoted the first half would be marketing; the second half is where a buyer should start asking questions.

There is another limitation in the comparison. Meituan's technical appendix marks some competing-model scores as sourced from public reports or official benchmark leaderboards, while other entries are self-run and described as indicative. The evaluation setups also differ by benchmark, with different agent harnesses, judges and sequence limits. The table is useful for identifying where Meituan believes its model is competitive, but it should not be read as a uniformly controlled ranking of every model shown.

The 1.6 trillion figure is now official

The parameter count was questioned during the preview period, when the same 1.6 trillion total and approximately 48 billion active parameters appeared in LongCat-2.0's model description. OrcaRouter subsequently reported that the figures attached to LongCat-2.5 had been carried over from the older model card.

Meituan's November 1 technical blog now explicitly identifies LongCat-2.5 as a 1.6-trillion-parameter model, with approximately 48 billion parameters activated per token. It also says the model builds on LongCat-2.0.

That settles what the company is claiming about the architecture, although it does not independently verify the parameter count. The important distinction is between an official specification and an externally inspectable model. With the weights still unavailable, outside researchers cannot yet independently inspect the architecture or reproduce the reported results from the model itself.

The more interesting technical question is what changed between versions. Meituan describes longer agent trajectories, large-scale reinforcement learning, failure-derived training tasks and native visual tokenization. Those are the improvements the company is asking developers to evaluate, rather than a newly announced increase in model scale.

What is actually available today

The preview has been usable since September 25 through Meituan's chat service and API (TechNode). The November 2 change log entry names compatibility with Claude Code, Codex and Hermes, and the platform carries setup guides for more harnesses, which means the realistic way to test LongCat-2.5 today is to point an existing agent setup at Meituan's endpoint and run it on your own workload.

GitHub and Hugging Face releases are still marked as forthcoming, with no publication date stated. Until the weights become available, developers can evaluate the model through Meituan's API, but cannot independently inspect the released architecture or reproduce the full evaluation using publicly available weights. The W4A4 optimization and domestic-chip performance claims also remain dependent on Meituan's technical disclosures.

Practical version: try the API on a long task with a known answer, and score it against the model you already use. Wait for the weights before treating the architecture as inspectable, and let someone other than Meituan run the table before treating its scores as settled.

Sources

  1. [1] Meituan / LongCat team — “LongCat-2.5 Released: Meituan's 1.6T Native Multimodal Agent Model,” November 1, 2026 (official technical release post: 1.6 trillion total parameters with roughly 48 billion activated per token, the same envelope as LongCat-2.0; training on longer agent trajectories and large-scale RL with fail-to-pass / pass-to-pass verification, failure-trajectory analysis and anti-reward-hacking hardening; native discrete visual tokens; benchmark summary including Terminal Bench v2.1 88.4 vs 70.8, BrowseComp 91.9, PaperBench 82.1, Program Bench 25.5 with Claude Opus 5 ahead by 11.5; evaluation note that asterisked competitor scores come from public reports, blogs or leaderboards and unmarked ones are self-run and indicative only; W4A4 INT4 inference on domestic accelerators, +43.3% prefill and +42.8% decode throughput vs W8A8 in Meituan's online stress test; “No open-weight release has been announced”; read first-hand)Read source
  2. [2] Meituan / LongCat team — “LongCat-2.5 Benchmarks: Full 33-Task Evaluation,” November 2026 (complete official benchmark table and methodology notes: Terminal Bench v2.1 88.4 vs 70.8; AutomationBench v1.0 65.1 vs 51.8; BrowseComp 91.9 vs 79.9; PaperBench 82.1 vs 45.2; 3DCodeBench (Text→3D) 55.4; Vision2Web 78.6; Program Bench 25.5 vs 37.0 for Claude Opus 5; asterisked scores taken from public technical reports, blog posts or official leaderboards, unmarked scores self-run and indicative only; read first-hand)Read source
  3. [3] LongCat API change log — September 25, 2026 entry: “LongCat-2.5-Preview Now Available” (new image understanding capability; coding; works with Claude Code, Hermes, OpenClaw, OpenCode and Kilo Code); November 2, 2026 entry: “LongCat-2.5 Now Available” (native multimodality, long-horizon agents; compatible with Claude Code, Codex, Hermes and other agent harnesses); read first-handRead source
  4. [4] TechNode — “Meituan releases LongCat-2.5, a 1.6-trillion-parameter native multimodal agent model,” November 2, 2026 (reporting context on the announcement, tracing to the same release: training emphasis, preview availability since September 25 via the chat service and API)Read source
  5. [5] AI Affairs — “Meituan Releases LongCat-2.5, a Native Multimodal Agent Model Built for Long-Horizon Tasks — Full Technical Report Published,” November 2, 2026 (reporting context; carried the preview-period OrcaRouter parameter dispute, and an AutomationBench pairing — 51.8 vs 15.8 — that Meituan's official table does not support: the official AutomationBench v1.0 figures are 65.1 for LongCat-2.5 and 51.8 for LongCat-2.0)Read source
  6. [6] Daily AI Brief — “Meituan Releases LongCat-2.5, a Native Multimodal Agent Model Built for Long-Horizon Tasks,” November 2, 2026 (reporting context on the announcement, tracing to the same release)Read source
  7. [7] LongCat team — LongCat-2.0 model page and release post, June 30, 2026 (the baseline model: 1.6 trillion total parameters, about 48 billion active, released as open weights under MIT; read first-hand)Read source