LongCat-2.0

Released June 30, 2026; open weights published to Hugging Face under the MIT license on July 4. LongCat-2.0 is a 1.6T-parameter open Mixture-of-Experts model (~48B active per token) with LongCat Sparse Attention and a native 1M-token context, trained end-to-end on Chinese-made accelerators. It brings near-frontier agentic-coding quality to open weights at a fraction of US-lab prices.

LongCat-2.0 strengths

• Open weights under a permissive MIT license
• Near-frontier agentic coding and tool use
• Native 1M-token context (128K max output)
• Very low launch pricing with free cached-input reads
• Efficient LongCat Sparse Attention (~48B of 1.6T active)

Pricing & context

Cost guide: a typical call of about 10K input + 2K output tokens costs roughly $0.005 at list prices. Note: $0.30 / $1.20 is a launch promotion — standard list pricing is $0.75 / $2.95 per 1M, with cached-input reads free.

When to choose LongCat-2.0

LongCat-2.0 is best for cost-efficient near-frontier agentic coding and open-weight, self-hosted deployments. If you want a frontier US-lab model instead, weigh it against DeepSeek V4 Pro or Kimi K2.7 Code.

LongCat-2.0 FAQ

How much does LongCat-2.0 cost?

LongCat-2.0 is priced at $0.30 per 1M input tokens and $1.20 per 1M output tokens as a launch promotion (standard list price $0.75 / $2.95), with a 1M-token context window and free cached-input reads. A typical 10K-input / 2K-output call costs roughly $0.005.

What is LongCat-2.0 best for?

LongCat-2.0 by Meituan is best for cost-efficient near-frontier agentic coding and open-weight, self-hosted deployments.

Is LongCat-2.0 open source?

Yes. Meituan released LongCat-2.0's weights on Hugging Face under the permissive MIT license, so you can self-host or fine-tune it.

How does LongCat-2.0 compare to DeepSeek V4 Pro?

Both are large open Mixture-of-Experts models. LongCat-2.0 (1.6T total / ~48B active, 1M context) launched at $0.30 input, undercutting DeepSeek V4 Pro's $0.435, though its $1.20 output is higher than DeepSeek's $0.87. The two trade blows on agentic-coding benchmarks.

Tools that use LongCat-2.0