Back to Blog
Open Source

Meituan's LongCat-2.0: A 1.6-Trillion-Parameter Coding Model Trained Entirely on Chinese Chips, Now MIT-Licensed

Meituan has officially open-sourced LongCat-2.0, a Mixture-of-Experts (MoE) coding model with a massive 1.6 trillion parameters. This is more than just another model release: it's believed to be the largest frontier AI model ever trained completely on domestic Chinese chips, with no Nvidia GPUs involved.

For the global AI industry, LongCat-2.0 sends a clear signal: despite chip export controls, China retains the capability to train models that compete with the world's best.

Model Specifications

ItemSpecification
DeveloperMeituan
ArchitectureSparse MoE
Total Parameters1.6 Trillion (1.6T)
Active Parameters~48 Billion (48B, dynamic range 33B-56B)
Context Window1M tokens
Training Data35 Trillion tokens
LicenseMIT (Fully Open-Source)
Training Hardware50,000 domestic AI chips (No Nvidia GPUs)
Communication LibraryHuawei HCCL
Special TechniquesLongCat Sparse Attention, 135B N-gram Embedding Module

Benchmark Performance

BenchmarkLongCat-2.0GPT-5.5Claude Opus 4.8
SWE-bench Pro59.558.669.2
Terminal-Bench 2.170.878.282.7

LongCat-2.0 achieved a score of 59.5 on SWE-bench Pro, surpassing GPT-5.5's 58.6. This is a landmark achievement: an open-source model trained entirely on domestic chips has beaten OpenAI's flagship model on a coding benchmark.

However, on broader agent benchmarks (FORTE, BrowseComp), LongCat-2.0 still trails Claude Opus 4.8, indicating closed-source flagships maintain an advantage in complex reasoning and multi-step tasks.

API Pricing

LongCat-2.0 is available on OpenRouter (previously listed anonymously as "Owl Alpha"):

Pricing TierInput Price (per 1M tokens)Output Price (per 1M tokens)
Standard$0.75$2.95
Promotional$0.30$1.20
Cache HitFree

Competitive Comparison:

ModelInput PriceOutput Price
LongCat-2.0$0.75$2.95
Kimi K2.7 Code$0.95$4.00
Claude Sonnet 5$2.00$10.00
GPT-5.5$5.00$30.00

LongCat-2.0's pricing is approximately one-seventh that of GPT-5.5. It's also about 20% cheaper than Kimi K2.7 Code, another open-source model.

Why This Matters

1. Independent Innovation Amid Chip Sanctions

Trained on 50,000 domestic AI chips using Huawei's HCCL library, without a single Nvidia GPU, LongCat-2.0 demonstrates:

  • China can train frontier-scale AI models under chip sanctions.
  • Domestic AI chip clusters have reached production-ready levels.
  • The software ecosystem (HCCL) is progressively replacing CUDA.

2. Fully Open-Source Under MIT License

Unlike many "open-weight" models with restrictions, LongCat-2.0 uses the MIT License—one of the most permissive open-source licenses. This means:

  • Commercial use without royalties.
  • Modification and redistribution are allowed.
  • Full local deployment with no usage limits.

3. Coding Capability Approaches the Frontier

A SWE-bench Pro score of 59.5 means LongCat-2.0 performs comparably to GPT-5.5 on real-world GitHub issue resolution tasks. For enterprises looking to deploy large-scale coding agents, it offers an extremely cost-effective option.

What This Means for Developers

Use CaseRecommended ModelRationale
Cost-sensitive large-scale coding agentsLongCat-2.0$0.75/M input, MIT license
Everyday coding assistanceKimi K2.7 CodeBetter Copilot integration & ecosystem
Complex reasoning & architecture designClaude Sonnet 5Strongest reasoning capabilities
Data-sensitive environmentsLongCat-2.0MIT license, fully local deployment
IDE-integrated workflowsGPT-4oMost mature Copilot ecosystem

Conclusion

The release of LongCat-2.0 represents more than just a new model—it signals a potential shift in the AI industry landscape.

Key takeaways:

  • A 1.6-trillion-parameter MoE model trained entirely on domestic chips, without Nvidia GPUs.
  • SWE-bench Pro 59.5 surpasses GPT-5.5, setting a new bar for open-source coding models.
  • MIT License offers maximum permissiveness for commercial use.
  • API pricing at $0.75/M input—roughly one-seventh the cost of GPT-5.5.
  • A major milestone in chip sovereignty, proving frontier models can be trained under sanctions.

For developers and enterprises, LongCat-2.0 presents a compelling option: near-frontier performance, extremely low cost, and full open-source freedom. In an increasingly diverse model landscape, cost and flexibility are becoming as critical as raw capability in the decision-making process.

Comments (0)

Share:XHatena

Post a Comment

Loading...