LongCat-2.0

LongCat-2.0 (June 2026) is Meituan's 1.6T-parameter open MoE (~48B active) built for agentic coding with a native 1M context — notable for being trained and served entirely on Chinese-made chips, under MIT.

Best for

• Agentic, long-horizon coding at frontier scale
• Native 1M-token context for large codebases
• Permissive MIT open weights for self-hosting
• Very cheap API with free context-cache hits

How it compares — and the India angle

• Edges GPT-5.5 on SWE-bench Pro (vendor-reported) at a fraction of the cost — appealing for cost-sensitive Indian builders.
• 1.6T total but only ~48B active per token, so inference is far cheaper than the size suggests.
• Benchmarks are vendor-reported and not yet independently verified — treat as indicative.