Meituan LongCat-2.0 Fully Open-Sourced Under MIT: A 1.6T MoE Making "Chinese AI Sovereignty" an Engineering Reality
The Event
On July 5, 2026, Meituan fully open-sourced LongCat-2.0 — MIT license, model weights plus inference code released, with no usage restrictions. This came within two weeks of the June 30 release: an upgrade from "public release" to "fully open source," marking another step forward for Chinese large models in openness.
Technical specifications:
- Architecture: MoE (Mixture-of-Experts), 1.6T total parameters, average 48B activated per token, dynamic range 33B–56B
- Context: 1M token native support
- Training hardware: reportedly the industry's first trillion-parameter model trained end-to-end (training and inference) on a 50,000-card domestic Chinese compute cluster
- Core technical innovations:
- LongCat Sparse Attention (LSA): a sparse attention mechanism for efficient 1M-token long-context processing
- Zero-Compute Experts: dynamic activation of 33B–56B parameters with zero wasted compute
- MOPD (Multi-Objective Pretrained Distillation): routes Agent / Reasoning / Interaction expert groups by task
- "MIT + 1.6T + 50K domestic cards" is the most persuasive evidence yet for Chinese AI sovereignty. Arguments questioning whether "Chinese AI truly has its own path" now face direct pushback
- SWE-bench Pro 59.5 > GPT-5.5 58.6 is a historic moment — the first stable inversion where a Chinese open-source flagship beats a US closed-source flagship on AI coding evaluation. Cursor, Anthropic, and OpenAI all have to take this seriously
- "AI coding toolchain ecosystem binding" went from a first case (Grok + Cursor) on 06-30 to a pattern (Meituan + Claude Code/OpenClaw/Hermes) on 07-05 — Chinese large models are becoming first-class citizens in the global AI coding toolchain
- A new battleground in the "AI deployment company" narrative: Together AI at an $11B valuation (07-03), Microsoft Frontier at $2.5B (07-03), and Meituan's full open-sourcing of LongCat-2.0 (07-05) — in the same week, majors treated "running models in customer scenarios" as more important than "the model itself." LongCat-2.0 pushes "AI deployment democratization" to the extreme via MIT full open-source + dual-platform deployment
- The specific model types, interconnect architecture, and training efficiency of the 50K-card domestic cluster are undisclosed, so outsiders cannot assess the real training time/cost behind "50K domestic cards"
- MIT license + commercial use means any company (regardless of nationality) can take the weights and build products for free. Where is the commercial moat? What is Meituan's next move?
- SWE-bench Pro 59.5 measures task completion rate, not code quality. Given that on 07-05's Senior SWE-Bench even Opus 4.8 scored only 24%, performance under a true "senior engineer standard" remains unknown
- Real-world NPU deployment performance: the NPU inference branch in the GitHub repo was just released; third-party benchmarks are pending
- https://x.com/Meituan_LongCat/status/2073768940078317713
- https://news.qq.com/rain/a/20260630A0420F00
- https://www.cnblogs.com/vibecodinghuanzhe/p/20991791
- https://aihot.virxact.com/items/cmr7vpphy007oslgnz4aotpx4
Benchmark results (official version, updated July 5):
| Benchmark | LongCat-2.0 | Note | |---|---|---| | Terminal-Bench 2.1 | 70.8 | Terminal tasks | | SWE-bench Pro | 59.5 | Above GPT-5.5's 58.6 | | SWE-bench Multilingual | 77.3 | Multilingual coding | | FORTE | 73.2 | Financial reasoning | | RWSearch | 78.8 | Real-world search | | BrowseComp | 79.9 | Browser control |
Native AI coding integration: Claude Code, OpenClaw, Hermes Agent — not a bolted-on afterthought, but protocol-level integration at launch.
Deployment flexibility: dual GPU + NPU platforms, already validated at scale on domestic clusters. Released simultaneously on HuggingFace, GitHub, and ModelScope.
Deep Analysis
The real significance here is not "yet another open-source model" but that it turns "Chinese AI sovereignty" from a political slogan into an engineering reality.
Three dimensions delivered in the same week:
1. Compute sovereignty: 1.6T MoE trained on a 50,000-card domestic cluster
The LongCat team's domestic compute exploration started from a thousand cards in 2023, spending three years on operator adaptation, communication optimization, and fault-tolerant scheduling. This time the claim is "industry first" — not "also uses it," but "the industry's first domestic cluster capable of running trillion-parameter workflows end-to-end." This means Chinese AI vendors now have concrete evidence of hardware independence, and follow-up models from DeepSeek, Alibaba, and ByteDance could take the same path.
2. Model sovereignty: 1M context + SOTA on 60+ benchmarks is not "good enough"
SWE-bench Pro 59.5 surpassing GPT-5.5 (58.6) is a representative signal in AI coding: a Chinese large model holding a stable lead over an OpenAI flagship on an internationally neutral platform for the first time. Terminal-Bench 2.1 at 70.8, FORTE at 73.2, BrowseComp at 79.9 represent SOTA in multiple niche tracks. This is not "good enough in Chinese" — it's SOTA on internationally recognized benchmarks.
3. Ecosystem sovereignty: native integration with Claude Code / OpenClaw / Hermes
The most crucial step: LongCat-2.0 was connected at the protocol level to mainstream AI coding toolchains at launch. This means Chinese large models are no longer "alternatives to foreign toolchains" but are "actively supported by top foreign toolchains." The "tri-party binding" of Cursor, Grok, and LongCat was a first case on 06-30; by 07-05 it had become a pattern.
The engineering reality of three stacked layers
"Sovereignty" has been overused. LongCat-2.0 offers a clear, quantifiable definition:
> Sovereignty = compute (domestic 50K cards) + model (1.6T MoE with 1M context) + ecosystem (native Claude Code/Agent protocol integration) — a closed triangle where none can be missing.
Why It Matters
Risks and Open Questions
Conclusion
Meituan's full open-sourcing of LongCat-2.0 is not "just another large model." It is the most persuasive engineering evidence in the 2026 Chinese AI sovereignty narrative — a closed triangle of compute, model, and ecosystem, with the MIT license ensuring this combination serves not only Meituan but the entire Chinese AI industry.
For the AI coding industry: the SWE-bench Pro first tier now includes a Chinese name — and it's open source. Toolchains like Cursor and Claude Code gain a SOTA option that does not depend on US hardware.
This is the answer to "can Chinese large models stand on their own?" Not "yes" or "no," but "it's already happening, with a landmark engineering milestone achieved."
---
Sources: