Bottom line: Meituan has released LongCat-2.5, a native multimodal Mixture-of-Experts model aimed at long-horizon agentic coding—not short chat demos. Per Meituan’s official LongCat-2.5 blog (dated 2026-11-1), the model has 1.6 trillion total parameters with about 48 billion activated per token, builds on LongCat-2.0, adds unified vision+text discrete tokens, and is available now through Meituan’s online chat and API at longcat.ai. Open weights on GitHub and Hugging Face are planned to follow. Treat published benchmark numbers as Meituan-reported (scores marked * for peers come from public reports), and do not overclaim versus Claude-Opus-5.
Secondary coverage: TechNode’s Nov 2, 2026 write-up summarizes the same launch for an English audience. This briefing complements our recent posts on DeepSeek’s open coding-agent prototype, GitHub Copilot cloud sessions and MCP connectors, Vercel agent-desktop accessibility-tree computer use, and Gemini 4 Carbon benchmarks — different stacks, same industry push toward longer agent runs.
What shipped: LongCat-2.5 in one page
LongCat-2.5 is Meituan’s next step after LongCat-2.0. The company positions it for tasks that need sustained agent execution: complex codebases, full-project builds, visual coding, creative design, and 3D. That is a different product story than “answer this coding question in one turn.” The blog stresses three pillars:
- Scale + sparsity: 1.6T total parameters, ~48B activated per token (MoE).
- Native multimodality: visual signals as multi-level discrete tokens so vision and text share one token pathway in the backbone.
- Long-horizon agent training: more long-horizon trajectories in long-sequence training, large-scale long-horizon RL, and converting real agent failure trajectories into reproducible training tasks.
Availability today is product-side: Meituan online chat plus API. Weights on GitHub/Hugging Face are promised afterward; as of this scan the public HF org still surfaces LongCat-2.0 as the home landing, so treat open weights as upcoming rather than already mirrored.
Meituan’s metrics dashboard also places LongCat-2.5 next to GPT-5.5, GPT-5.6-Sol, Claude-Opus-4.8, and Claude-Opus-5 in charts, with a fuller table later in the post. For readers who only need a shipping decision: hosted multimodal coding agent now; open weights later; numbers competitive on several agent benches with clear gaps still visible against Opus-5 on harder terminal suites.
How it differs from chat demos (long-horizon agents)
Most public coding-model demos still look like autocomplete or single-file edits. Long-horizon agentic coding means multi-step loops: explore a repo, run tools, fail, recover, keep context across a long trajectory, and ship something closer to a project than a snippet. Meituan’s blog describes scaling an end-to-end agentic RL pipeline across data, environments, algorithms, and harnesses—including synthesizing tasks from GitHub PRs/issues with fail-to-pass and pass-to-pass tests, asynchronous partial rollouts, staleness controls, and anti–reward-hacking measures (disabled network in task envs, git history removed, LLM judges auditing exploits).
For developers evaluating models, the practical question is not “did it write a pretty README?” but whether the harness + model can stay coherent across long tool-use sessions. That is the same problem space we covered for DeepSeek’s Router–Worker research prototype, Copilot’s cloud sandboxes with snapshot resume, and Vercel’s accessibility-tree computer-use path—each attacking long-running agent work from a different angle.
Mixed training across scaffolds is another detail worth noting: Meituan says the model adapts to similar tasks and harnesses during RL, mixing benchmark-aligned tasks and scaffolds so gains reinforce one another. In product terms, that is Meituan’s bet that one backbone can generalize across terminal agents, browse agents, and visual coding harnesses instead of shipping a zoo of fine-tunes.
Architecture highlights (brief)
Meituan frames visual primitive modeling as discrete tokens rather than a bolt-on vision encoder in the hot path. Per the blog:
- RVQ + SAE: Residual Vector Quantization on native-resolution features from a Semantic-and-Aligned Encoder (as in LongCat-Next), producing multi-level vocabulary tokens for images.
- Offline visual preprocessing: a fixed tokenizer so visual token IDs can be precomputed/cached; the backbone consumes pretokenized sequences, which simplifies long-sequence packing and load balancing.
- Joint visual embeddings: codebook embeddings learned with the backbone; Meituan cites preliminary t-SNE evidence of interwoven visual/textual embeddings on a smaller MoE (~3B activated on average).
- Teacher–student transfer: retains LongCat-2.0 expert groups (agent, reasoning, interaction) with online student sampling and per-cut prefixes / Tree Attention to control distribution shift and compute.
Infrastructure notes matter if you care about domestic-accelerator deployment: long-context training optimizations (supernode-aware parallelism, dual-shard FSDP, train–inference numerical consistency) and an efficient W4A4 inference path for the 1.6T MoE. Versus a W8A8 INT8 baseline in Meituan’s online stress tests: prefill throughput +43.3%, decode +42.77%, TTFT −15.91%, TPOT −19.72%, and GMM kernel about 1.73×. Those are vendor figures for their stack—useful as directional efficiency claims, not portable scores for your GPU cloud.
Selected Meituan-reported metrics
The official table compares LongCat-2.5 with LongCat-2.0 and several peers (GPT-5.5, GPT-5.6-Sol, Claude-Opus-4.8, Claude-Opus-5, GLM-5.3, kimi-k3, DeepSeek-V4.1-Flash). Peer scores marked with * are from public reports/leaderboards; otherwise self-run and indicative. Below is a compact subset for coding, agent, and visual tasks:
| Benchmark | LongCat-2.5 | LongCat-2.0 | Claude-Opus-5 (note) |
|---|---|---|---|
| Terminal Bench v2.1 | 88.4 | 70.8 | 89.1* |
| Terminal Bench v4.0 | 38.5 | 3.0 | 53.9* |
| PaperBench | 82.1 | 45.2 | 79.1 |
| BrowseComp | 91.9 | 79.9 | 90.8* |
| Vision2Web | 78.6 | — | 80.2* |
| DeepSWE v1.1 | 64.6 | 9.7 | 68.8* |
| MCP Atlas | 83.1 | 61.8 | 85.8* |
Reading the table carefully: LongCat-2.5 is a large jump over 2.0 on Terminal Bench v2.1 (88.4 vs 70.8) and sits near the top cluster, but Claude-Opus-5’s public * score is still slightly higher at 89.1, and Opus-5 leads more clearly on Terminal Bench v4.0 (53.9* vs 38.5). PaperBench and BrowseComp look strong for LongCat-2.5 on Meituan’s numbers. Vision2Web at 78.6 trails Opus-5’s 80.2*. Do not market “beats Opus-5 across the board”—the honest story is “competitive MoE coding/agent model with native multimodality and a big lift from 2.0.”
How to try LongCat-2.5 today
- Chat: Open longcat.ai and use Meituan’s online chat interface for interactive trials (visual coding demos are highlighted on the blog).
- API: Use the API platform linked from the official blog / longcat.ai for programmatic access—confirm rate limits, thinking-mode flags, and context caps in the current docs before production wiring.
- Weights (planned): Watch Meituan’s GitHub and Hugging Face orgs for LongCat-2.5 checkpoints. Until they land, do not assume local/offline deploy parity with the hosted chat/API.
- Eval hygiene: If you re-run Terminal Bench / SWE-style harnesses, match harness versions and settings Meituan documents (e.g., thinking enabled, temperature 1.0, top_p 0.95 for many of their reported runs).
One more practical distinction: LongCat-2.5’s showcases (fluid simulation tools, 3D exhibitions, city dioramas, business analysis packs) are framed as end-to-end agent outputs spanning software development and office workflows. That marketing is useful as a capability map—visual + coding + document work—but your own acceptance tests should still be repo- and harness-specific. A fluid-sim demo does not automatically transfer to your monorepo’s flaky integration suite.
How this sits next to DeepSeek, Copilot, Vercel, and Gemini posts
DeepSeek’s MIT-licensed dsh-superagent-prototype is an open research coding-agent scaffold (Router–Worker, observe/auto auth)—architecture and openness first. GitHub Copilot’s connected cloud sessions emphasize portable sandboxes, snapshot resume, and verified MCP connectors billed against Actions entitlement. Vercel agent-desktop attacks computer use via accessibility trees for coding agents. Gemini 4 Carbon is Google’s frontier model narrative with public benchmark positioning. LongCat-2.5 is Meituan’s hosted multimodal MoE play for long-horizon engineering agents, with open weights promised later. Teams should compare on harness fit, multimodal needs, data residency / domestic-accelerator constraints, and whether they need open weights now or can start on API.
If you already run Copilot cloud sandboxes or a Vercel computer-use desktop loop, treat LongCat-2.5 as a candidate model backend for the same long jobs rather than a replacement IDE. The interesting experiment in November 2026 is cross-vendor: same harness, same repo fixture, same max steps—swap only the model/API—and publish the delta with harness versions attached.
Admin / developer checklist
- Pilot on hosted chat/API first for a multi-hour repo task (tests failing → fix → PR draft) before betting on weights that are not public yet.
- Separate multimodal jobs (Vision2Web-style UI rebuilds, 3D/code visuals) from text-only SWE loops so you can see where native vision tokens help.
- Record harness versions when comparing against Terminal Bench / PaperBench / BrowseComp so Meituan-reported numbers stay apples-to-apples.
- Watch HF/GitHub for LongCat-2.5 weight drops; plan quantization and serving (W4A4 claims are on Meituan’s domestic accelerators—validate on your hardware).
- Policy & data: confirm what code you may send to Meituan’s API under your org’s outbound AI policy.
- Don’t overclaim in marketing vs Claude-Opus-5; cite Meituan’s table and the * footnotes.
- Budget for long context: long-horizon rollouts are token-heavy; measure cost per successful task, not per message.
- Keep a failure log the way Meituan describes training—your own agent failures are the best seeds for eval sets once weights arrive.
FAQ: Meituan LongCat-2.5
What is Meituan LongCat-2.5?
LongCat-2.5 is Meituan’s native multimodal MoE model (1.6T total parameters, ~48B activated per token) built for long-horizon agentic tasks such as complex codebase work, full-project builds, visual coding, creative design, and 3D. It builds on LongCat-2.0 with stronger visual understanding and agentic coding training.
When was LongCat-2.5 released, and where can I try it?
Meituan’s official blog is dated 2026-11-1; TechNode covered the launch on 2026-11-02. You can try it now via Meituan’s online chat and API at longcat.ai. GitHub and Hugging Face weight releases are planned to follow.
How is LongCat-2.5 multimodal?
It represents visual signals as multi-level discrete tokens (Residual Vector Quantization on Semantic-and-Aligned Encoder features) so the backbone processes vision and text in a unified token format, with offline visual preprocessing and jointly learned visual embeddings.
How do LongCat-2.5 Terminal Bench scores compare to Claude-Opus-5?
On Meituan’s table, LongCat-2.5 scores 88.4 on Terminal Bench v2.1 versus LongCat-2.0’s 70.8 and Claude-Opus-5’s 89.1* (public report). On Terminal Bench v4.0, LongCat-2.5 scores 38.5 versus Opus-5’s 53.9*. It is competitive on v2.1 but does not lead Opus-5 across both suites.
What W4A4 inference gains did Meituan claim?
Versus W8A8 INT8 in Meituan’s online stress tests on their domestic-accelerator stack: prefill throughput +43.3%, decode +42.77%, TTFT −15.91%, TPOT −19.72%, and GMM kernel ~1.73×. These are vendor infrastructure results, not universal GPU-cloud guarantees.
Are LongCat-2.5 weights on Hugging Face yet?
As of this briefing, Meituan says GitHub/Hugging Face releases are planned to follow the chat/API launch. Public HF org pages still centered on LongCat-2.0 at scan time—verify the org before assuming 2.5 checkpoints are downloadable.
Bottom line: LongCat-2.5 is Meituan’s Nov 2026 multimodal MoE for long-horizon agentic coding—1.6T / ~48B active, discrete vision tokens, heavy long-horizon RL, chat+API live today, open weights next. Use Meituan’s own table (with * footnotes) when comparing to Opus-5 and peers; pilot on longcat.ai for real multi-step repo work; and watch HF/GitHub for when local deploy becomes possible.
Leave a Reply
You must be logged in to post a comment.