LongCat 2.0: Meituan's 1.6T MIT-Licensed Model for Agents

The assumption that frontier-scale open models come only from dedicated AI labs stopped holding some time ago. LongCat 2.0 comes from Meituan, the Chinese food-delivery and local-services company, and it is a 1.6-trillion-parameter mixture-of-experts model released under the MIT licence.

The weights went up on Hugging Face in the first week of July 2026, with FP8 and INT8 checkpoints alongside the main release, and OpenRouter lists it from 20 July. Meituan's card gives no formal release date, so date comparisons should use the listing you actually call.

From LongCat-Flash to 2.0

Meituan's previous flagship line, LongCat-Flash, launched in 2025 as a 560B MoE whose active parameter count varied per token, between 18.6B and 31.3B and averaging about 27B. LongCat 2.0 nearly triples the total to 1.6T, with roughly 48B active per token. The Hugging Face listing's automatic parameter count shows a larger figure, around 1.8T, so check which number a comparison table is using.

The jump in active parameters matters more than the jump in total. Around 48B active means each token costs meaningfully more compute than LongCat-Flash did, which shows up in serving cost and latency. Active vs total parameters covers why that is the number to read first.

Architecture: what is new beyond the size

Three pieces of the design stand out on the model card:

• LongCat Sparse Attention (LSA), a custom sparse attention scheme aimed at 1M-token workloads. Meituan says the model saw hundreds of billions of tokens of 1M-context data in training.
• An N-gram Embedding of about 135B parameters, carried over from LongCat-Flash-Lite. It adds capacity in the embedding layer rather than in the experts.
• A three-step multi-token prediction module, used for speculative decoding to speed up generation.

The card also states pre-training on more than 35T tokens, and says training and deployment were built entirely on AI ASIC superpods. It does not name the chips, so treat any claim about whose silicon that is as unconfirmed.

OpenRouter lists a context window of about 1M tokens and up to 262,144 output tokens. A trained 1M window and a reliable one are different things; every 1M-context model compared covers how to tell them apart.

The benchmark table, read against the right reference

Meituan's card reports scores measured in-house under a unified harness, with competitors' figures cited from their own reports where marked, which covers most of the Opus 4.8 coding rows. Against Claude Opus 4.8, the strongest model in the table:

• Terminal-Bench 2.1 — 70.8 against 78.9
• SWE-bench Pro — 59.5 against 69.2
• SWE-bench Multilingual — 77.3 against 84.8
• BrowseComp — 79.9 against 84.3
• GPQA-diamond — 88.9 against 92.4
• IFEval — 90.0 against 86.0

It beats Opus 4.8 on IFEval and on IMO-AnswerBench (81.8 against 75.3), but Gemini 3.1 Pro and GPT-5.5 both score higher on IFEval, so those rows are not a lead over the field. The fairer reading is that it trails the strongest closed model on agentic coding by roughly seven to ten points. Against the Gemini 3.1 Pro and GPT-5.5 figures in the same table it is ahead on SWE-bench Pro, level on SWE-bench Multilingual and Terminal-Bench 2.1 against Gemini, and three points behind GPT-5.5 on Terminal-Bench. That is a strong result for an MIT-licensed model, and it is not frontier parity.

Same shape as DeepSeek V4 Pro

The closest comparison is DeepSeek V4 Pro, which is almost the same size: 1.6T total with 49B active, a 1M context and an MIT licence. On paper the two are interchangeable, which makes the differences the deciding factors.

Price is one. LongCat 2.0 is listed on OpenRouter at $0.30 per million input tokens and $1.20 output at the time of writing; DeepSeek V4 Pro has been around $0.44 input and $0.87 output. LongCat is cheaper on input-heavy work such as long-context reading, and DeepSeek is cheaper on output-heavy work such as long generations or verbose reasoning.

The other is what each was tuned for. Meituan stresses integration with agent harnesses such as Claude Code, OpenClaw and Hermes, and its benchmark selection leans towards terminal and browsing agents. Run both on your own loop; with identical licences and near-identical shapes, nothing on paper decides it.

Thinking modes and what they cost

The card documents three ways to run it: thinking on, thinking on while keeping the full reasoning content across turns (which Meituan says performs better), and thinking off for token efficiency. Keeping reasoning content means re-sending it every turn, so in a long agent session it compounds input cost quickly. Start with thinking on without preserved reasoning, and only switch to preserved reasoning if you measure a difference on your tasks.

Self-hosting at 1.6T

The weights are on Hugging Face and ModelScope, with separate FP8 and INT8 checkpoints. Meituan's card points to SGLang for GPU serving and an SGLang fork for NPUs, without a GPU count; at 1.6T parameters, even FP8 weights need well over a terabyte of accelerator memory. That is a multi-node-class deployment for most teams, and the MIT licence does not make the hardware any cheaper. For the full picture, MIT-licensed models compared covers what the licence does and does not let you build.

Pick LongCat 2.0 when you need permissive weights at frontier-adjacent scale and your workload is agentic and input-heavy. Pick something smaller when you need to run it yourself on ordinary hardware, and something closed when the last ten points on SWE-bench Pro are worth paying for.