Meituan LongCat-2.0 Fully Open Sourced Under MIT License With 1.6T MoE Weights

Meituan announced the full open source release of LongCat-2.0 under an MIT license. The move includes complete model weights for the 1.6 trillion parameter MoE architecture and all inference code.

The announcement came through the official LongCat account on X. It marks one of the largest parameter counts made available under a permissive license to date.

The release supplies both GPU and NPU deployment paths that have already been validated on large domestic clusters.

This decision places immediate pressure on closed model providers who have kept similar scale systems private. It also gives developers direct access to a long-context agent model without usage caps or API keys.

Meituan states that LongCat-2.0 activates roughly 48 billion parameters per token while supporting a 1 million token context window. It ships with three specialized expert groups for agent, reasoning, and interaction tasks.

The company also introduced LongCat Sparse Attention and Zero-Compute Experts to reduce wasted computation during long sequence processing.

Benchmark results reported with the release show Terminal-Bench 2.1 at 70.8, SWE-bench Pro at 59.5, and SWE-bench Multilingual at 77.3. Additional scores include FORTE at 73.2, RWSearch at 78.8, and BrowseComp at 79.9.

These numbers place the model ahead of GPT-5.5 on the SWE-bench Pro metric by a margin of 0.9 points.

The code supports native integration with Claude Code, OpenClaw, and Hermes Agent frameworks. This integration lowers the barrier for teams that already rely on those agent toolchains.

Direct Access To Training Scale Changes Evaluation Norms

Teams can now download the full 1.6T weights and run local experiments that previously required commercial API access. This removes a major cost barrier for academic groups and smaller startups.

The MIT license permits both commercial use and modification. No additional restrictions on derivative models were listed in the release notes.

Prior open releases at this parameter count stayed limited to research checkpoints or required custom licenses. The LongCat-2.0 package removes that friction.

Smaller Closed Labs Face Immediate Benchmark Pressure

Closed providers that charge for agent capabilities now compete against a freely available model that reports higher scores on several SWE-bench variants. The comparison is direct because the weights and evaluation harness are public.

Larger labs may respond by accelerating their own open releases or by tightening terms on existing APIs. Smaller labs without comparable training infrastructure face a steeper gap.

The release does not include training data or the full pre-training recipe. This leaves reproduction of the original results dependent on Meituan's published inference stack alone.

Technical Choices Target Long Context And Sparse Activation

LongCat-2.0 uses a Mixture-of-Experts design that routes each token through only a subset of experts. The active count per token stays near 48 billion even though total parameters reach 1.6 trillion.

Zero-Compute Experts allow the system to skip inactive experts without incurring compute cost. This mechanism widens the range of active parameters from 33 billion to 56 billion depending on task demand.

MOPD routing directs queries toward one of three expert pools labeled Agent, Reasoning, or Interaction. The router aims to match task type before generation begins.

LongCat Sparse Attention replaces standard attention for sequences beyond 128k tokens. The method reduces memory growth while preserving retrieval accuracy on the 1 million token context window.

These design decisions focus on deployment cost rather than raw capability scaling. They reflect a practical trade-off between model size and inference economics.

Verification Gaps Remain On Third-Party Hardware

All reported benchmarks and cluster tests come from Meituan's own evaluation environment. Independent runs on widely available GPU hardware have not yet appeared in public forums.

The NPU support path is documented only for specific domestic accelerators. Porting to other NPU vendors will require additional engineering work by the community.

The release includes a reference implementation but does not guarantee performance parity across every hardware combination. Early adopters will need to validate throughput and memory usage themselves.

Next Milestones To Track Over The Next Quarter

Watch for independent benchmark reproductions on A100 and H100 clusters within the next 30 days. Consistent scores in those environments would strengthen the original claims.

Observe whether agent frameworks such as OpenClaw publish official integration guides that reference the released weights. Official support would accelerate adoption.

Track any responses from closed model vendors in their next model cards or API pricing updates. Direct comparisons in vendor materials would confirm competitive impact.

Monitor community fine-tunes on the Hugging Face hub. Derivative models that exceed the base scores on SWE-bench would indicate the weights are useful beyond the original checkpoint.

These signals will clarify whether the open release shifts practical agent development away from paid APIs or remains limited to research settings.