Meituan launches 1.6T multimodal LongCat-2.5, open weights still pending

Meituan’s LongCat team released LongCat-2.5 on November 1, 2026, together with a full technical report, and the model became available on the LongCat API platform the next day. LongCat-2.5 is a natively multimodal mixture-of-experts model with 1.6 trillion total parameters and about 48 billion active per token. It is built for long-running agent work, and its Hugging Face and GitHub weights are listed as “coming soon,” not yet published.

The headline change is not size. The parameter count, context window and output length match the September preview. What Meituan changed is the training target and the vision support. Its own numbers show large jumps on multi-step agent benchmarks compared with LongCat-2.0.

What LongCat-2.5 changes over LongCat-2.0

LongCat-2.5 uses the same basic design as its predecessor: 1.6T total parameters, roughly 48B active per token, a 1M-token context and 128K output. One coverage report calls this “the same envelope as the September preview”. A mixture-of-experts (MoE) model routes each token through only a subset of its parameters. That is why only about 48B of the 1.6T are active at any step.

According to the same report, the shift is in purpose. LongCat-2.0 was built to write code, while 2.5 is built to finish projects across a codebase, a terminal, a browser, a spreadsheet or a design canvas. It can also take images as input. Meituan’s technical report says long-horizon agentic reinforcement learning, not pretraining scale, produced the gains. Reinforcement learning (RL) here means training the model on multi-step tasks in real environments.

An API directory listing summarises the vendor’s stated focus as agentic coding and long-running engineering tasks, visual coding such as creative design and 3D work, and image understanding inside agent workflows. The vendor also says the model works across agent harnesses such as Claude Code, Codex and Hermes.

Spec	LongCat-2.5 (as reported)
Architecture	Mixture of experts
Total / active parameters	1.6T / about 48B per token
Context window	1,048,576 tokens
Output length	128K; an API directory lists 131,072 max output and a 262,144 max_tokens cap
Input / output	Text and images in, text out
Reasoning control	thinking parameter, on or off
Open weights	“Coming soon” on GitHub and Hugging Face; licence not confirmed

How the native vision works

“Native multimodal” here means Meituan did not attach a separate vision encoder to a text model. Images are converted into multi-level discrete tokens using residual vector quantization (RVQ) over native-resolution features from a sparse autoencoder (SAE). This approach comes from Meituan’s LongCat-Next line. As a result, text and vision share one token stream and one training objective. RVQ represents an image as a stack of codes, coarse first and finer after. An SAE is a network that learns compact features from data.

The visual tokens are precomputed and cached offline, so the vision encoder never enters the training graph. According to coverage of the report, a probe model with about 3B active parameters showed visual and text embeddings interwoven rather than separated into clusters. The directory listing describes the same idea in simpler terms: variable-resolution images become discrete tokens.

In practice, Meituan presents this as “visual coding.” The model reads a screenshot, chart, user interface or 3D viewport inside the same reasoning loop that writes the code. The API returns text only. LongCat-2.5 can read images but does not generate them.

Benchmarks: Meituan’s numbers, and where it trails

All scores below come from Meituan’s announcement and technical report. No independent leaderboard has published LongCat-2.5 results in these sources. In Meituan’s comparison against LongCat-2.0, the 2.5 model moves:

• Terminal Bench v2.1: 70.8 to 88.4
• Terminal Bench v4.0: 3.0 to 38.5
• DeepSWE v1.1: 9.7 to 64.6
• PaperBench: 45.2 to 82.1
• BrowseComp: 79.9 to 91.9

Other reported results include AutomationBench 51.8 and MathVision 96.5. On Terminal Bench v4.0, coverage of the launch notes that LongCat-2.5’s 38.5 is ahead of GPT-5.6-Sol’s 37.3 in Meituan’s comparison set.

The same report shows clear gaps. Claude Opus 5 leads Terminal Bench v4.0 at 53.9, about 15 points ahead of LongCat-2.5. DeepSeek V4.1-Flash leads on DeepSWE v1.1 and Terminal Bench v2.1. Pure vision results sit mid-pack, with VideoMME at 85.2 against Kimi K3’s 90.0. AILog summarises Meituan’s own conclusion this way: competitive breadth with an edge in visual coding, but not yet dominant on the hardest engineering tasks.

Independent evidence on the LongCat family is still thin. As of November 9, the third-party comparison site BenchLM listed LongCat-2.0 as “not ranked” in its coding and agentic categories. For a pairing with Xiaomi’s MiMo-V2.6-Flash, it reported zero shared benchmark results, so it does not name a winner. The sources contain no BenchLM entry for LongCat-2.5. As one launch write-up puts it, benchmarks remain claims until reproduced in your own environment.

Where to run it: hosted API now, weights later

LongCat-2.5 became available on Meituan’s LongCat API platform on November 2. Third-party coverage, cited by AILog, confirms the November 2 date. Meituan also offers an online chat experience alongside the API. An API directory lists four routes to the model: LongCat itself, Vercel AI Gateway, ZenMux and NanoGPT. It uses the model ID meituan/longcat-2.5 on an OpenAI-style /chat/completions endpoint.

The documented controls matter for anyone integrating the model:

• A thinking parameter turns step-by-step reasoning on or off, using {"type":"enabled"} or {"type":"disabled"}.
• Temperature is fixed at 1.0 with thinking on and 0.7 with it off.
• max_tokens and thinking_budget are each capped at 262,144.

The output limit is reported inconsistently. AILog gives 128K output. The directory lists a 131,072-token maximum output but also, citing LongCat’s API parameters page, a 262,144-token max_tokens cap. Builders should check the limit against their own calls.

The open-weight question

This is the main gap for self-hosters. Meituan’s launch page shows Hugging Face and GitHub weights as “coming soon,” so LongCat-2.5’s open-weight status and licence are not confirmed. AILog goes further and says no open-weight release has been announced. It calls this a break from the family default, since Flash-Chat, Flash-Thinking, LongCat-2.0 (MIT licence), Next and Video-Avatar all shipped weights. GoDirX reads the same “coming soon” note more optimistically, as weights being on the way.

Meituan’s Hugging Face organization page still describes itself as the “Home of LongCat-2.0”. The LongCat-2.0 repository includes a chat template with tool-calling support and an enable_thinking switch. For now, LongCat-2.0 remains the newest LongCat general model that can be downloaded and run locally. LongCat-2.5 should be treated as a hosted service, with the usual dependencies on latency, data residency and pricing.

Serving efficiency: 4-bit inference on domestic chips

Meituan says it tuned both long-context training and inference for domestic Chinese accelerators, specifically cards with under 80GB of memory each. For inference, it uses W4A4 quantization, which stores expert weights and activations at 4 bits. It recovers accuracy with learnable equivalent transformations in the style of FlatQuant. The company says the quantized model matches the unquantized one on every evaluated downstream task.

Against 8-bit (W8A8) serving in online stress tests, Meituan reports:

• Prefill throughput up 43.3%
• Decode throughput up 42.77%
• Time to first token (TTFT) down 15.91%

On the training side, Meituan reports that its FSDP (fully sharded data parallel) setup delivers 86% higher per-card throughput than a DDP (distributed data parallel) baseline on long-sequence supervised fine-tuning. These are vendor figures from Meituan’s own infrastructure. The sources do not include independent checks.

What it means for builders

For teams building coding agents, LongCat-2.5 adds another large hosted model aimed at multi-step work. It can read screenshots and UI state in the same loop that writes code. The vendor says it is compatible with harnesses such as Claude Code and Codex, which could make it easy to trial. Its 1M-token context leaves room for large codebases.

Several limits apply:

• Evidence is self-reported. Every benchmark comes from Meituan, and the model trails Claude Opus 5 on the hardest terminal benchmark in Meituan’s own table.
• No weights yet. Self-hosting, fine-tuning and private deployment depend on a release that is listed as coming soon but has no date or licence.
• Scale. If the weights do arrive, a 1.6T model still needs very large infrastructure and engineering effort to run.
• Text-only output. Pipelines that need generated images need a separate model.

Meituan’s Hugging Face page also points to a related project, LongCat-DeepResearch. It pairs a research-tuned LongCat model with a harness for evidence-grounded reports. Meituan says the system beats the evaluated deep-research offerings from ChatGPT, Claude and Gemini on DeepResearchBench, DeepResearchBench II and ResearchRubrics. Meituan states that these model improvements are intended for the next general-purpose LongCat release.

What to watch next

• The weights. Whether LongCat-2.5 appears on Hugging Face and GitHub, and under what licence. LongCat-2.0 used MIT.
• Independent scores. Third-party runs of Terminal Bench, DeepSWE and BrowseComp, which leaderboards such as BenchLM currently lack for the LongCat family.
• Output limits. A consistent public figure for maximum output length.
• The next general model. Whether the LongCat-DeepResearch improvements show up in the next LongCat release, as Meituan says it intends.