A 1.6-trillion-parameter Mixture-of-Experts model trained entirely on Chinese chips, released under MIT license, scoring nearly 60% on SWE-Bench Pro. LongCat-2.0 is the open-source coding model nobody saw coming.
What Is LongCat-2.0?
LongCat-2.0 is Meituan's answer to the question: can Chinese labs produce world-class open coding models without NVIDIA hardware? Released June 29, 2026, under an MIT license, it's a 1.6-trillion-parameter Mixture-of-Experts (MoE) model purpose-built for software engineering tasks.
The key differentiator: it was trained entirely on domestic Chinese chips. No H100s, no A100s — just whatever Meituan could get its hands on domestically. And it still posts numbers that rival Western frontier models on coding benchmarks.
Benchmark Performance
| Benchmark | LongCat-2.0 | Claude Sonnet 5 | DeepSeek V4-Pro |
|---|---|---|---|
| SWE-Bench Pro | 59.5% | ~62% | ~58% |
| Terminal-Bench | 70.8 | ~68 | ~66 |
| Context Window | Long-context | 200K | ~1M tokens |
Those SWE-Bench Pro numbers put LongCat-2.0 firmly in frontier territory for open-source models. Only Claude Sonnet 5 and GPT-5.6 Sol clearly beat it, and both are proprietary with API-only access.
Why This Matters
1. Open Source at Frontier Scale
1.6 trillion parameters under MIT license is rare. Most models at this scale — GPT-5.6, Gemini 3.5 Pro — are locked behind APIs. DeepSeek V4-Pro is open-weight at similar scale, but LongCat-2.0 specifically targets the coding use case with specialized architecture.
2. Trained Without NVIDIA
The entire training run used domestic Chinese accelerators. This proves that export controls haven't stopped Chinese labs from producing competitive models — they've just accelerated investment in alternative hardware ecosystems.
3. Coding-Specific Architecture
Unlike general-purpose LLMs, LongCat-2.0 was optimized for: - Multi-file code generation and refactoring - Long-context code understanding (entire repositories) - Debugging and test generation - Terminal and shell workflow integration
Practical Use Cases
For developers: Download the weights, run locally or on cloud GPU instances. MIT license means commercial use is fully permitted — no API costs, no rate limits, no vendor lock-in.
For enterprises: Self-host LongCat-2.0 behind your own firewall for code assistance. This matters for organizations with strict data residency requirements or those who can't send proprietary code to third-party APIs.
For researchers: The MoE architecture at this scale is a goldmine for study. How did Meituan achieve 59.5% on SWE-Bench Pro without frontier Western hardware? The open weights let the community dig in.
Hardware Requirements
Running a 1.6T MoE locally isn't trivial. At 4-bit quantization, expect to need roughly 800GB+ of VRAM. That's multi-GPU territory: - 8× H100 80GB: feasible but expensive - 4× H200 141GB: more comfortable - Cloud A100/H100 instances: practical for most teams
The MoE architecture helps — not all 1.6T parameters are active during inference. But this is still a heavyweight model designed for serious infrastructure, not a laptop toy.
How It Fits the July 2026 Landscape
LongCat-2.0 fills an important niche in the open-source hierarchy:
- DeepSeek V4-Pro (1.6T MoE): Best general-purpose open model
- LongCat-2.0 (1.6T MoE): Best open coding-specific model
- GLM-5.2 (Zhipu AI): Best mid-size open general model
- Kimi K2.7 Code (Moonshot AI): Best lightweight open coding model
For teams evaluating open-source coding models right now, the choice is really between LongCat-2.0 for maximum performance and Kimi K2.7 Code for a lighter footprint.
Limitations
- General reasoning: LongCat-2.0 is coding-specialized. For general QA, creative writing, or analysis, DeepSeek V4-Pro or GLM-5.2 will serve better.
- English-heavy training: While multilingual, the coding benchmarks skew toward English-language codebases. Non-English performance may vary.
- Deployment complexity: This is not a model you spin up in five minutes. Plan for real infrastructure work.
- Newness: Released June 29 — the community is still finding edge cases and failure modes.
Bottom Line
LongCat-2.0 is a landmark release. It proves that open-source coding models can reach near-frontier performance, that Chinese labs can train at scale without NVIDIA, and that the MIT license continues to dominate as the preferred license for serious open AI work.
If your team needs a self-hosted coding model and has the infrastructure to run it, LongCat-2.0 deserves a shortlist spot alongside DeepSeek V4-Pro.