Skip to main content
MindStudio
Pricing
BlogAbout
My Workspace
LongCat 2.5Meituan AI modelLongCat MoE

LongCat 2.5: Meituan's 1.6T Model Looks Good on Paper, Shaky in Practice

LongCat 2.5 is Meituan's 1.6T parameter MoE model for agentic coding. Hands-on tests show gaps between its benchmarks and real output.

Edited by Luis Chavez-Mattos, Director of Product RSS
LongCat 2.5: Meituan's 1.6T Model Looks Good on Paper, Shaky in Practice

What is LongCat 2.5?

LongCat 2.5 is a mixture-of-experts language model from Meituan, the Chinese food delivery and local services company, released under an MIT license. It carries 1.6 trillion total parameters with roughly 48 billion active per token, and it handles both text and images as input. Meituan built it mainly for agentic work: long coding sessions inside large codebases, multi-step project builds, and tool-using tasks rather than short chat exchanges. The model is distributed across 208 safetensors shards on Hugging Face plus separate multimodal weight files, reflecting its scale.

TL;DR

  • LongCat 2.5 is a 1.6 trillion parameter MoE model from Meituan with about 48 billion active parameters per token, licensed under MIT.
  • The model takes text and images natively at variable resolution, using a vision tokenizer that turns each image into a list of discrete tokens without resizing or cropping.
  • Meituan trained the model’s agentic coding ability using real GitHub pull requests and issues, built into sandboxed test environments with pass/fail checks and an error judge that penalizes cheating.
  • Benchmark tables show a large jump over LongCat 2, with some scores beating Claude Opus on tests like BrowseComp and 3D code generation, though it trails on terminal-bench and other coding benchmarks.
  • Meituan itself flags that unstarred scores were measured in-house and should be treated as indicative rather than verified.
  • Independent hands-on testing found a gap between the curated demo gallery and actual output, including a failed simulation task and a multilingual naming task that defaulted to literal translation instead of cultural grounding.
  • Access remains the biggest practical barrier: the model is too large to run locally for most people, and third-party API access has been inconsistent, including through free offers tied to OpenCode.
Self-host
Remy.
On-prem or in your private cloud.
‹ ›
https://claims.northwind.internalYour servers. Your domain.
Open claims1,902
Approved today214
Flagged17
Avg. review3.2h
CLM-20417 · M. Alvarez$4,120Approved
CLM-20418 · D. Okafor$880In review
CLM-20419 · S. Lindqvist$12,300In review

How big is LongCat 2.5 and what does the architecture look like?

LongCat 2.5 sits in the upper tier of open-weight model sizes, with 1.6 trillion total parameters spread across a mixture-of-experts design. Only about 48 billion parameters activate for any given token, which is the standard MoE trade-off: a large total capacity with a much smaller active compute cost per step. The weights are published in 208 safetensors shards, plus additional multimodal parameter files, which gives a sense of just how much storage and VRAM a full local deployment would require.

The model’s vision handling is notable on the technical side. Rather than resizing every image to a fixed resolution, LongCat 2.5 uses a vision tokenizer that converts each image into a list of discrete token IDs, where larger images simply produce longer token lists. A newspaper page, a handwritten equation, a toy photo, and a close-up of an animal can all be fed in at their native size and dimensions. A detokenizer can reconstruct the image from those same tokens, which means images and text end up represented as the same basic currency inside the model. Meituan says these image tokens can be precomputed and cached ahead of training, which simplifies the training pipeline.

How was LongCat 2.5 trained for agentic coding?

The documentation describes a training loop built specifically around software engineering tasks. Real GitHub pull requests and issues get converted into working environments, each with tests designed to fail before a fix is applied and pass afterward, plus regression tests that should keep passing. The model then attempts these tasks inside a sandboxed environment with no web search access. Rewards come from a combination of the test results and an automated judge that checks for shortcuts or reward hacking, with penalties for repetitive reasoning or answers that run unnecessarily long.

This is a sound approach for teaching a model real debugging and coding judgment rather than pattern-matching on static benchmark questions. It mirrors how other frontier labs have approached agentic training: grounding reward signals in actual test execution rather than human preference alone.

What do the benchmark claims actually show?

Meituan’s published comparison table shows a sizable jump from LongCat 2 to LongCat 2.5 on agentic and coding benchmarks, with some scores moving from clearly behind the field to roughly competitive. Against closed frontier models, the picture is mixed. On tests like BrowseComp and a 3D code generation benchmark, LongCat 2.5 is reported slightly ahead of Claude Opus in Meituan’s own table. On others, including terminal-bench and program-bench style tests, it falls clearly behind.

The more important detail is a disclosure buried in the benchmark notes: scores without an asterisk were measured in-house by Meituan and are labeled as indicative only, not independently verified. That is a meaningful caveat. It means a chunk of the headline numbers come from the same team that built the model, run under conditions nobody outside Meituan has reproduced. Treating the benchmark table as a starting point rather than a verdict is the more reasonable way to read it.

The enterprise platform for building software with AI.
Secure, compliant, auditable apps, built to scale.

How does LongCat 2.5 perform in real-world testing?

Hands-on testing against the model’s own promotional material revealed gaps worth noting. Meituan’s technical blog features a showcase of demos, generated end-to-end by LongCat 2.5, spanning categories like office productivity tools, Chinese-style visual designs, a real-time fluid simulation, a 3D exhibition space, and a city diorama. These are presented as polished, finished outputs.

A direct test asked the model to take an image of a doner kebab stand and simulate it as a single web page. The result was a loosely rendered shape described as a brown leaf-shaped piece of meat on a metal pole next to a grill block and an unclear gray cutout, with an odd label underneath. It did not resemble a doner kebab simulation and did not match the polish of the curated demo gallery. This was a single run, so it is not a definitive verdict on the model’s capability, but it illustrates the distance between marketing material and unscripted output.

A second test asked LongCat 2.5, run through OpenCode, to generate a name for a hypothetical new political party in 83 different languages, with the requirement that each name be unique and ideally grounded in local culture. The model did correctly use the right script for every language on the list, which is a real achievement for multilingual coverage. But nearly every name turned out to be a literal translation of the phrase “unique party” rather than a culturally distinct or semantically creative name. Some names were nearly identical across closely related languages, such as Hindi and Bhojpuri, and in at least one case (Nepali) the model left its own correction notes visible in the output instead of producing a clean final answer.

Is LongCat 2.5 worth trying right now?

The architecture and training methodology are genuinely well documented. Native variable-resolution image tokenization, an MIT license, and a training loop built on real GitHub issues with sandboxed test verification are all sensible design choices for an agentic coding model, and the technical write-up explains them in real detail.

The practical problem is access. At 1.6 trillion total parameters, running LongCat 2.5 locally is out of reach for almost anyone without serious multi-GPU infrastructure. Third-party API access has also been inconsistent: free access offers tied to tools like OpenCode have proven difficult to sustain even a short session with, and there is no simple hosted chat interface to test the model casually. Combined with the gap between the curated demo gallery and actual test runs, the most reasonable conclusion is to treat LongCat 2.5 as a model worth watching rather than one with a settled reputation. Anyone evaluating it for real work should run their own tasks against it rather than relying on the benchmark table or the demo showcase alone.

Frequently Asked Questions

What company makes LongCat 2.5?

Meituan, a Chinese company best known for food delivery and local services, develops the LongCat model family.

How many parameters does LongCat 2.5 have?

NOV17Tuesday
MindStudioRemy

A big announcement for MindStudio builders

  • 11:30am PT · 2:30pm ET · 1 hour
  • Live on Zoom · with Q&A
Hosted byDmitry ShapiroCEO & Co-Founder

It has 1.6 trillion total parameters in a mixture-of-experts architecture, with approximately 48 billion active per token.

Can I run LongCat 2.5 locally?

Technically yes, since the weights are open under an MIT license, but the model’s size (208+ safetensors shards) makes local deployment impractical without substantial multi-GPU hardware.

Are LongCat 2.5’s benchmark scores verified?

Meituan states that benchmark scores without an asterisk in its comparison table were measured in-house and should be treated as indicative rather than independently confirmed.

What is LongCat 2.5 designed to do best?

It is built primarily for agentic coding tasks, such as working inside large codebases, using real GitHub pull request and issue data combined with sandboxed test verification during training.

Editorial standards

Presented by MindStudio

No spam. Unsubscribe anytime.