Are LongCat-2.0 weights actually downloadable yet?

LongCat-2.0's Hugging Face repository at meituan-longcat/LongCat-2.0 currently shows only the README and license file; the model card states 'Model weights coming soon.' Meituan has launched the chat interface at longcat.ai and posted the inference code on GitHub, but the full 1.6T checkpoint upload is pending as of June 30, 2026.

What's new in LongCat-2.0's architecture?

LongCat-2.0 introduces two pieces beyond standard MoE: a 'LongCat Sparse Attention' mechanism that keeps the 1M-token context affordable at inference time, and a 135B-parameter N-gram Embedding module that adds capacity in a sparse dimension orthogonal to the expert layout. Activation is dynamic, ranging from 33B to 56B parameters per query.

Was LongCat-2.0 really trained without NVIDIA chips?

LongCat-2.0 was pretrained on a 50,000-card cluster of domestic Chinese AI ASIC superpods, per the official tech blog. VentureBeat's coverage confirms the run used no Nvidia A100s or H100s and no AMD MI300X chips — Meituan's entire training stack is Chinese-manufactured silicon, making this the largest open-weight model trained on a domestic compute base.

How does LongCat-2.0 perform on coding and agent benchmarks?

LongCat-2.0 reports 59.5 on SWE-bench Pro, 70.8 on Terminal-Bench 2.1, and 78.3 on SWE-bench Multilingual. On agent search it hits 79.9 on BrowseComp and 78.8 on RWSearch. Foundational reasoning numbers include 80.0 on IMO-AnswerBench, 87.9 on GPQA-Diamond, and 63.7 on MRCR v2 (8 needle).

What license is LongCat-2.0 under and can companies fine-tune it?

LongCat-2.0 is released under the MIT license, one of the most permissive licenses available. Once the weights ship on Hugging Face, companies will be free to download, fine-tune, redistribute, and deploy them commercially without paying royalties or seeking permission from Meituan.

Meituan · 2026-06-30 · major

LongCat-2.0 — Meituan's 1.6T open-source MoE for agentic coding

Meituan released LongCat-2.0, a 1.6T-parameter open-source mixture-of-experts model with ~48B active per token. It scores 59.5 on SWE-bench Pro and 70.8 on Terminal-Bench 2.1, trained entirely on Chinese AI ASICs under MIT license.

Meituan's first frontier-tier open-source coding model — 1.6T parameters with 48B active, trained without a single NVIDIA chip.

Key specs

Swe bench pro	59.5
Total parameters	1.6T (48B active)

Quick facts

Maker	Meituan
License	MIT
Architecture	Mixture-of-Experts, dynamic 33B–56B active
Context window	1M tokens
Terminal-Bench 2.1	70.8
Training stack	50K+ domestic AI ASICs, 35T tokens
Availability	longcat.ai chat + HF (weights coming soon)

What is it?

LongCat-2.0 packs 1.6 trillion total parameters into a mixture-of-experts layout that activates roughly 48 billion per token. Meituan, the Chinese super-app group, released it under the MIT license through Hugging Face and GitHub on June 30, 2026, alongside a chat interface at longcat.ai.

How does it work?

Three new pieces drive the design: 'LongCat Sparse Attention' for the 1M-token window, a 135B-parameter N-gram Embedding module that scales capacity orthogonal to the MoE expert layout, and dynamic activation that ranges from 33B to 56B parameters per query. Pretraining ran over 50,000 domestic AI ASICs for millions of accelerator-hours across 35 trillion tokens.

Why does it matter?

Open-weight coding agents now have a Chinese-chip-trained option in the same league as leading closed models on SWE-bench Pro and Terminal-Bench 2.1. The MIT license lets teams self-host and fine-tune without restriction, the 1M-token context handles whole-repo agent work, and the entirely domestic training stack — no NVIDIA H100s, no AMD MI300X — makes LongCat-2.0 a strategic choice for groups facing US export controls.

Who is it for?

agentic-coding teams, Chinese-stack adopters, export-controlled groups

Frequently asked questions

Are LongCat-2.0 weights actually downloadable yet?: LongCat-2.0's Hugging Face repository at meituan-longcat/LongCat-2.0 currently shows only the README and license file; the model card states 'Model weights coming soon.' Meituan has launched the chat interface at longcat.ai and posted the inference code on GitHub, but the full 1.6T checkpoint upload is pending as of June 30, 2026.
What's new in LongCat-2.0's architecture?: LongCat-2.0 introduces two pieces beyond standard MoE: a 'LongCat Sparse Attention' mechanism that keeps the 1M-token context affordable at inference time, and a 135B-parameter N-gram Embedding module that adds capacity in a sparse dimension orthogonal to the expert layout. Activation is dynamic, ranging from 33B to 56B parameters per query.
Was LongCat-2.0 really trained without NVIDIA chips?: LongCat-2.0 was pretrained on a 50,000-card cluster of domestic Chinese AI ASIC superpods, per the official tech blog. VentureBeat's coverage confirms the run used no Nvidia A100s or H100s and no AMD MI300X chips — Meituan's entire training stack is Chinese-manufactured silicon, making this the largest open-weight model trained on a domestic compute base.
How does LongCat-2.0 perform on coding and agent benchmarks?: LongCat-2.0 reports 59.5 on SWE-bench Pro, 70.8 on Terminal-Bench 2.1, and 78.3 on SWE-bench Multilingual. On agent search it hits 79.9 on BrowseComp and 78.8 on RWSearch. Foundational reasoning numbers include 80.0 on IMO-AnswerBench, 87.9 on GPQA-Diamond, and 63.7 on MRCR v2 (8 needle).
What license is LongCat-2.0 under and can companies fine-tune it?: LongCat-2.0 is released under the MIT license, one of the most permissive licenses available. Once the weights ship on Hugging Face, companies will be free to download, fine-tune, redistribute, and deploy them commercially without paying royalties or seeking permission from Meituan.

Try it

https://longcat.ai