
LongCat-2.0
Meituan's open-weights 1.6T-parameter MoE language model for coding and agentic tasks, with 1M-token context.
LongCat-2.0 is an MoE model from Meituan's LongCat team with 1.6 trillion total parameters and about 48 billion active per token, open-sourced in a 2026-06-30 launch post with weights under the MIT License. On the LongCat API it costs a limited-time $0.30 per million uncached input tokens and $1.20 per million output tokens.
About LongCat-2.0
Meituan's LongCat team calls LongCat-2.0 a substantial step up from its earlier LongCat models. The launch post says the full training run and large-scale deployment are built entirely on AI ASIC superpods, with pretraining on over 50K AI ASICs across more than 35 trillion tokens and no rollbacks or irrecoverable loss spikes, and that hundreds of billions of tokens of 1M-context data went into strengthening it on long-horizon tasks.
The design builds on LongCat-Flash: LongCat Sparse Attention, an evolution of DeepSeek Sparse Attention with a lighter indexer, speeds up long-context processing, an N-gram Embedding module adds 135B parameters, and a 3-step Multi-Token Prediction module speeds up speculative decoding. Post-training merges agent, reasoning and interaction expert groups, and the team says the model is deeply integrated with harnesses such as Claude Code, OpenClaw and Hermes.
In LongCat's evaluation table, where LongCat-2.0's own scores are measured in-house, it scores 70.8 on Terminal-Bench 2.1, 59.5 on SWE-bench Pro, 77.3 on SWE-bench Multilingual and 79.9 on BrowseComp. Its SWE-bench Pro score is above the reported 54.2 for Gemini 3.1 Pro and 58.6 for GPT-5.5, while Claude Opus 4.8's reported scores are higher on all three coding benchmarks, at 78.9, 69.2 and 84.8.
You can chat with LongCat-2.0 on longcat.ai, and the model card says it can be deployed on both GPU and NPU platforms, with SGLang guides for each. The LongCat API Platform serves it in OpenAI and Anthropic API formats with a 1M-token context window and up to 128K output tokens, and also lists a LongCat-2.5-Preview model with the same limits.
Sources: longcat.chat/docs, /blog/longcat-2.0, huggingface.co/meituan-longcat/LongCat-2.0, longcat.ai/platform/docs/pricing/longcat-2.0 and longcat.ai/platform/docs, read 2026-10-27
LongCat-2.0 pros and cons
Pros
- The weights are released under the MIT License, and the model card says nothing in it alters or restricts those terms.
- At the limited-time API prices, cached input costs $0.006 per million tokens, against $0.30 for uncached input.
- The API takes requests in OpenAI format at https://api.longcat.ai/openai and in Anthropic format at https://api.longcat.ai/anthropic, so OpenAI and Anthropic SDK code can point at it.
- The chat template can switch thinking mode on or off, and the model card labels thinking off as better for token efficiency.
Cons
- LongCat-2.0's scores in the launch table are LongCat's in-house measurements, and the figures marked * for other models are external, reported metrics.
- The API prices are a limited-time discount, and LongCat says they are subject to change.
- LongCat's service agreement describes the LongCat service as meant for personal study, research, daily life and entertainment, and bars commercial use of the service or its outputs without Meituan's permission.
- The model card says the model has not been designed or comprehensively evaluated for every possible downstream application, and asks developers to assess accuracy, safety and fairness before sensitive or high-risk deployments.
Tagged
- Web-based
LongCat-2.0 alternatives
The closest alternatives to LongCat-2.0.
- 01→DeepSeek-V4-Flash-0731
Official release of DeepSeek-V4-Flash, with stronger agentic skills and open weights under the MIT License.
- 02→Kimi K3
Moonshot AI's 2.8T-parameter model with native vision and a 1M-token context for long-horizon coding.
- 03→Hy4 preview
Tencent's open-sourced 770B model with 49B active parameters and a 1M-token context, for long-horizon work.
Related reading
Featured on Vibedonalds
Own LongCat-2.0? Add this badge to your site to show you’re listed — and link back to your profile here.
<a href="https://vibedonalds.com/tools/longcat-2-0" target="_blank" rel="noopener nofollow">
<img src="https://vibedonalds.com/badge/featured-on-vibedonalds.svg" alt="LongCat-2.0 — Featured on Vibedonalds" width="240" height="60" loading="lazy" />
</a>Frequently asked questions
- What are the alternatives to LongCat-2.0?
- The closest alternatives to LongCat-2.0 are DeepSeek-V4-Flash-0731, Kimi K3, Hy4 preview.
- What are the downsides of LongCat-2.0?
- LongCat-2.0's scores in the launch table are LongCat's in-house measurements, and the figures marked * for other models are external, reported metrics. The API prices are a limited-time discount, and LongCat says they are subject to change. LongCat's service agreement describes the LongCat service as meant for personal study, research, daily life and entertainment, and bars commercial use of the service or its outputs without Meituan's permission. The model card says the model has not been designed or comprehensively evaluated for every possible downstream application, and asks developers to assess accuracy, safety and fairness before sensitive or high-risk deployments.
- What is LongCat-2.0?
- An MoE language model from Meituan's LongCat team with 1.6 trillion total parameters, about 48 billion of them activated per token, trained and deployed on AI ASIC superpods. LongCat pitches it at coding and agentic tasks, with training on 1M-context data for long-horizon work.
- Is LongCat-2.0 free and open source?
- LongCat says it is open sourcing the model, and the model card says the weights are released under the MIT License. The LongCat API is paid per token through Pay-As-You-Go, and the model card points to longcat.ai for chatting with the model.
- How much does the LongCat-2.0 API cost?
- The pricing page lists limited-time discounted prices per million tokens: $0.30 for uncached input, $0.006 for cached input and $1.20 for output. Prices are subject to change, and the LongCat API Platform and your billing records are authoritative.
- How does LongCat-2.0 do on benchmarks?
- LongCat's table gives it 70.8 on Terminal-Bench 2.1, 59.5 on SWE-bench Pro, 77.3 on SWE-bench Multilingual, 79.9 on BrowseComp, 90.0 on IFEval and 88.9 on GPQA-diamond. Its own scores are measured in-house; values marked * for other models are external, reported figures.
- Can I use LongCat-2.0 in Claude Code?
- LongCat says the model is deeply integrated with harnesses such as Claude Code, OpenClaw and Hermes, and it ran Terminal-Bench 2.1 and the SWE-bench series via Claude Code. The API Platform accepts requests in Anthropic format at https://api.longcat.ai/anthropic.