Meituan publishes LongCat-2.5 weights, opening a self-hosted route for visual coding

Developers can now run and modify Meituan’s visual coding model outside its hosted service. The company released LongCat-2.5’s weights on November 9, 2026, under the MIT license, one week after introducing the model. That opens a commercial reuse path for software that combines screenshot and video understanding with multi-step coding—but operating it requires the full checkpoint.

The checkpoint is available on Hugging Face. Its license permits commercial use, modification and redistribution, subject to notice requirements. Those permissions apply to the downloadable weights; buying access to Meituan’s hosted service remains a separate decision.

Meituan describes a model that processes visual information alongside text, allowing it to inspect screenshots, work through repositories and write or revise software across multiple steps. Its API documentation accepts images embedded in requests or supplied by URL. The documented video interface requires a URL—a practical constraint for teams using internal screen recordings.

Stronger than its predecessor, not a blanket rival win

Meituan’s clearest benchmark gains are against LongCat-2.0. The company reports improvements across terminal tasks, software engineering and questions about codebases. These are company-published results, rather than independent confirmation of how reliably the model will handle a team’s own software.

Meituan’s reported gains over LongCat-2.0

Terminal Bench v2.1

70.8→88.4

score

Meituan reports 70.8 for LongCat-2.0 and 88.4 for LongCat-2.5.

DeepSWE v1.1

9.7→64.6

score

Meituan reports 9.7 for LongCat-2.0 and 64.6 for LongCat-2.5.

SWE Atlas

33.1→58.3

score

On codebase question-answering, Meituan reports 33.1 for LongCat-2.0 and 58.3 for LongCat-2.5.

Comparisons with competitors are more uneven. In Meituan’s table, LongCat-2.5 comes within one point of GPT-5.6 Sol and Claude Opus 5 on Terminal Bench v2.1. On version 4.0, however, it scores 38.5 against Opus 5’s 53.9. DeepSeek-V4.1-Flash leads those four models on DeepSWE; LongCat leads on 3DCodeBench, while Opus 5 leads on Vision2Web.

The proprietary reference models also had successors available before LongCat-2.5’s introduction. GPT-6.1 Sol and Claude Opus 5.5 post higher Terminal-Bench 4.0 leaderboard scores, but use different testing setups. That prevents a clean head-to-head conclusion: the older-model near tie does not establish parity with the newer systems.

Hosted prices and the cost of independence

Meituan lists limited-time API rates at the following prices per million tokens—the units a model processes or generates:

• Uncached input: $0.30.
• Cached input: $0.006.
• Output: $1.20.

At those rates, 200,000 uncached input tokens and 100,000 output tokens cost $0.18. The same fixed volumes cost $1.40 with GPT-6.1 Sol and $2.80 with Opus 5.5 at their listed rates. Those are token-price comparisons, not equal-cost task comparisons: reasoning, repeated tool calls and second attempts can change consumption.

Self-hosting involves a different calculation. Meituan describes 1.6 trillion total parameters, with about 48 billion active per token. The model routes work through selected groups of parameters, known as experts, rather than using all of them each time. But the smaller active count does not remove the need to make the full checkpoint available.

There is also an unresolved size discrepancy: Hugging Face’s repository metadata displays 1.8T, while Meituan’s technical description says 1.6T. At two bytes per parameter, the latter would imply roughly 3.2 TB for weights alone. That is illustrative arithmetic, not the actual checkpoint footprint or a complete serving-memory requirement.

Meituan points operators to an SGLang-FluentLLM deployment guide for neural processing units. Its technical blog describes four-bit expert weights and activations on Chinese accelerators. Those optimizations describe Meituan’s tested setup, not a guarantee that another hardware fleet will preserve the same visual and tool-use behavior.