Here is a new article based on the provided post content:
—
# Meituan Unveils LongCat-2.0: A 1.6T-Parameter Open MoE Model Built for Agentic Coding
Meituan has introduced **LongCat-2.0**, a next-generation open large language model built on a Mixture-of-Experts (MoE) architecture. Designed specifically for agentic coding workflows, LongCat-2.0 combines massive scale with efficiency-focused innovations, enabling strong code understanding, generation, and execution in complex, multi-step tasks.
## A Trillion-Scale Model That Prioritizes Practical Efficiency
With **1.6 trillion total parameters** and approximately **48 billion parameters activated per token**, LongCat-2.0 represents a significant step forward in open-source MoE models. Unlike conventional dense models, MoE architectures activate only a subset of experts per token, reducing computational cost while maintaining high capacity.
The model is the successor to LongCat-Flash (560B), which was released in 2025. LongCat-2.0 was architected around one primary objective: **reliable and efficient agentic coding**, even when operating on non-Nvidia hardware.
## Key Capabilities
– **Native 1-million-token context window**: Enables whole-repository reasoning and multi-step terminal tasks without costly summarization.
– **Fully domestic training and inference**: Ran on **domestic AI ASIC superpods**, avoiding reliance on Nvidia GPUs.
– **Stable training at scale**: Pretrained on over **35 trillion tokens** with no rollbacks or irrecoverable loss spikes.
## Architectural Innovations That Cut Cost
What makes LongCat-2.0 remarkable is not just its size, but how efficiently it scales. Four core design choices make a 1.6T-parameter model practical:
1. **Zero-computation experts**
Simple tokens (such as punctuation) skip heavy computation entirely, routing through a zero-computation expert. A learned PID controller balances the mix of simple and complex tokens, keeping average activated capacity between **33B and 56B parameters**.
2. **LongCat Sparse Attention (LSA)**
Traditional attention scales quadratically with context. LSA selects only the most relevant tokens using three orthogonal indexing strategies:
– Streaming-aware indexing for contiguous memory access
– Cross-layer attention reuse
– Hierarchical, coarse-to-fine filtering
This enables the 1M-token window without a memory wall.
3. **N-gram embedding module**
A 135-billion-parameter embedding module captures dense local token relationships orthogonal to the MoE experts, reducing memory I/O during decoding.
4. **Post-training with MOPD**
A dedicated post-training pipeline fuses three teacher expert groups covering **Agent, Reasoning, and Interaction**, producing a unified model tuned for real-world workflows.
## Benchmark Results
Meituan reports that LongCat-2.0 performs competitively on agentic coding benchmarks:
| Benchmark | Score | Focus |
|———|——|——|
| SWE-bench Pro | **59.5** | Real-world software engineering |
| Terminal-Bench 2.1 | **70.8** | Shell execution and error recovery |
| SWE-bench Multilingual | **77.3** | Cross-language repository tasks |
Meituan claims LongCat-2.0 edges out GPT-5.5 on SWE-bench Pro and compares favorably to Google’s Gemini 3.1 Pro in software engineering tasks, though broader general-agent benchmarks still show room for improvement.
## Model Comparison at a Glance
| Feature | LongCat-2.0 | LongCat-Flash |
|——–|————-|—————|
| Total parameters | 1.6T | 560B |
| Active parameters/token | ~48B (33B–56B) | ~27B |
| Context window | 1M tokens (native) | 128K |
| Attention | LongCat Sparse Attention | Multi-head Latent Attention |
| Hardware | Domestic AI ASICs | H800 GPUs |
| License | MIT | MIT |
| Weights availability | Coming soon | Open |
## Practical Use Cases
LongCat-2.0 is optimized for **agent-based software development**, not casual chat. Ideal workloads include:
– **Whole-repository debugging and refactoring** across dozens of files
– **Multi-step terminal tasks** with error recovery via agent loops
– **Cross-language migration** in polyglot repositories
– **Large-scale codebase transformations** with coordinated edits
These patterns work with standard agent harnesses, lowering the adoption barrier for development teams.
## Access and Pricing
The model is available via the **LongCat API Platform**, with both OpenAI- and Anthropic-compatible endpoints. It is also supported on **OpenRouter** and in frameworks such as Claude Code, OpenClaw, OpenCode, and Codex.
– **Input price**: ~$0.75 per million tokens (launch promotion: $0.30)
– **Output price**: ~$2.95 per million tokens (launch promotion: $1.20)
– Cached context reads are free
Note: Local self-hosting is not yet available, as model weights are still pending public release.
## Final Thoughts
LongCat-2.0 demonstrates that extreme-scale MoE models do not have to be synonymous with extreme cost. By combining sparse attention, expert routing, and domestic hardware, Meituan delivers a trillion-parameter model that is both powerful and practical for real-world agentic coding tasks.
—
**Source**: *Marktechpost*, “Meituan Releases LongCat-2.0: A 1.6T Parameter Open MoE Model With Native 1M Context and LongCat Sparse Attention”, July 5, 2026. [Original article](https://www.marktechpost.com/2026/07/05/meituan-releases-longcat-2-0-a-1-6t-parameter-open-moe-model-with-native-1m-context-and-longcat-sparse-attention/)
