LongCat-2.5 explained: Meituan’s 1.6T agentic model.

LongCat-2.5 is the newest model from Meituan’s LongCat team, and it is aimed squarely at long-horizon agentic coding. The LongCat team says it has 1.6 trillion total parameters with about 48 billion activated per token, sees images natively alongside text, and is built for complex codebases, full-project builds, visual coding and 3D tasks. It went live on the LongCat API platform on November 2, 2026, at a limited-time price well below most frontier models. Here is what LongCat-2.5 is, what’s new, how to try it, and what people are already building with it on the LongCat-2.5 page.

What is LongCat-2.5?

LongCat’s announcement, dated November 1, 2026, describes LongCat-2.5 as “a native multimodal model for long-horizon agentic tasks.” It is a mixture-of-experts (MoE) model: of its 1.6 trillion parameters, roughly 48 billion are used for each token, which is how very large models keep inference costs manageable. It builds on LongCat-2.0, and the team says the upgrade is mainly about two things: stronger native visual understanding and stronger agentic coding.

TechNode, citing IT Home, reports that Meituan released the model on November 2 and that it is available through Meituan’s online chat service and API platform, with GitHub and Hugging Face releases planned to follow. The weights were not public at the time of writing.

What’s new in LongCat-2.5

Long-horizon agentic coding

The LongCat team says it added more long-horizon trajectories to long-sequence training and used large-scale reinforcement learning to improve end-to-end engineering work, from complex codebase development to building whole projects. It also says it turns real-world agent failures into reproducible, verifiable training tasks, so gaps it finds in the model become new training data. Coding tasks are synthesized from filtered GitHub pull requests and linked issues, then checked with fail-to-pass and pass-to-pass tests.

Native multimodality

Instead of bolting a separate vision encoder onto a language model, LongCat-2.5 represents images as multi-level discrete tokens, so text and visual inputs flow through the same backbone in one token format. The team says combining this with agentic coding gives it strong visual coding performance and support for creative design and 3D tasks.

Works with existing agent harnesses

LongCat’s API changelog says LongCat-2.5 is compatible with Claude Code, Codex, Hermes and other agent harnesses. Many of the announcement’s own coding scores were run in a Claude Code harness, which suggests that is a natural place to start.

Built on domestic accelerators

The announcement says long-context training and W4A4 (4-bit weights and activations) inference were optimized for domestic AI accelerators. LongCat reports that its 4-bit setup raised prefill throughput by 43.3% and decode throughput by 42.77% versus an 8-bit baseline in its own online stress tests.

LongCat-2.5 benchmarks

LongCat publishes a long evaluation table comparing LongCat-2.5 with GPT-5.5, GPT-5.6-Sol, Claude Opus 4.8, Claude Opus 5, GLM-5.3, Kimi K3 and DeepSeek-V4.1-Flash. These are LongCat’s own numbers, and the team notes that scores it ran itself for other models are indicative only. A few highlights from that table:

• Terminal Bench v2.1: 88.4 for LongCat-2.5, against 89.1 listed for Claude Opus 5 and 88.3 for Kimi K3.
• Terminal Bench v4.0: 38.5, up from 3.0 for LongCat-2.0; the table lists 53.9 for Claude Opus 5.
• BrowseComp: 91.9, the highest figure in LongCat’s table for that benchmark.
• DeepSWE v1.1: 64.6, below several of the models it is compared with, including 74.2 for DeepSeek-V4.1-Flash.
• 3DCodeBench Text-to-3D: 55.4, the top score in the table, in line with the team’s focus on visual and 3D work.

The honest read is that LongCat positions 2.5 as close to the frontier on agentic and browsing tasks, ahead on some visual and 3D tests, and still behind the strongest models on several hard coding benchmarks.

LongCat-2.5 pricing and context window

According to LongCat’s API docs, LongCat-2.5 has a 1M-token context window and can output up to 128K tokens in a single request. Pay-as-you-go pricing is listed as a limited-time discount:

• Uncached input: $0.30 per million tokens (¥2).
• Cached input: $0.006 per million tokens (¥0.04).
• Output: $1.20 per million tokens (¥8).

LongCat notes that prices are subject to change and that the prices shown on the LongCat API Platform and in your billing records are authoritative.

How to try LongCat-2.5

1. Chat with it in the browser at LongCat Chat↗ (opens in a new tab), which is where early testers have been running it.
2. For the API, register on the LongCat API Platform and create an API key on its API Keys page.
3. Call it with the model name LongCat-2.5 through the OpenAI-compatible endpoint (https://api.longcat.chat/openai) or the Anthropic-compatible endpoint (https://api.longcat.chat/anthropic).
4. Thinking can be switched on or off per request; LongCat’s docs say temperature is fixed at 1.0 with thinking on and 0.7 with it off.
5. To use it as a coding agent, point a compatible harness such as Claude Code or Codex at the Anthropic- or OpenAI-style endpoint.

Who is LongCat-2.5 for?

Based on what LongCat says the model is tuned for, it fits developers who want a low-cost model for long agentic coding sessions, teams that need a very long context window for big codebases, and creators doing visual coding, interactive web demos and 3D scenes. LongCat’s own showcase includes a real-time fluid simulation webpage, 3D digital exhibitions and city dioramas, alongside office tasks like business analysis and reports.

What people are building with LongCat-2.5

The first LongCat-2.5 build on the Atlas is a take on the classic pelican-riding-a-bicycle test, generated directly in the LongCat web chat: see Pelican riding a bicycle by LongCat-2.5. It is an early sign of how creators are probing the model’s visual coding, and we’ll add more builds to the LongCat-2.5 model page as they appear. To compare, browse what people have made with Claude Opus 5 and Kimi K3, two of the models in LongCat’s benchmark table.