Meituan LongCat-2.0 Fully Open-Sourced Under MIT: A 1.6T MoE Model Trained on 50K Domestic Accelerators
*English translation of a Chinese forum post. Claims and benchmarks below are as reported by the original author.*
The Event
On July 5, 2026, Meituan officially fully open-sourced LongCat-2.0 — MIT license, model weights + inference code released, no usage restrictions. This came just two weeks after the model's June 30 launch: an upgrade from "public release" to "full open source," marking another step forward for Chinese large models in terms of openness.
Technical specifications:
- Architecture: MoE (Mixture of Experts), 1.6T total parameters, average 48B activated per token, dynamic range 33B-56B
- Context: native 1M token support
- Training hardware: reportedly the first trillion-parameter model trained end-to-end (training and inference) on a 50,000-chip domestic compute cluster
- Core technical innovations:
- LongCat Sparse Attention (LSA): sparse attention mechanism for efficient 1M-token long-context processing
- Zero-Compute Experts: dynamically activates 33B-56B parameters with zero wasted compute
- MOPD (Multi-Objective Pretrained Distillation): routes across three expert groups — Agent / Reasoning / Interaction — per task
- "MIT + 1.6T + 50K domestic chips" is the most convincing evidence combination for China's AI sovereignty narrative
- 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
- AI coding toolchain ecosystem binding went from a first case on 06-30 (Grok + Cursor) to a pattern on 07-05 (Meituan + Claude Code/OpenClaw/Hermes) — Chinese large models are becoming "first-class citizens" of the global AI coding toolchain
- A new main battleground for the "AI deployment company" narrative: Together AI's $11B valuation (07-03), Microsoft Frontier $2.5B (07-03), and Meituan's full open-source release (07-05) all landed in the same week — major players now treat "running models in customer scenarios" as more important than the models themselves. LongCat-2.0 pushes "AI deployment democratization" to the extreme via MIT full open source + dual-platform deployment
- The specific accelerator model, interconnect architecture, and training efficiency of the 50K-chip cluster are undisclosed, so outsiders cannot verify real training time/costs
- MIT license + commercial use = any company (of any 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 the 07-05 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 on GitHub was just released; third-party validation is 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 | Notes | |---|---|---| | 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 integrations: Claude Code, OpenClaw, Hermes Agent — not a "late-stage hookup," but protocol-level integration at launch.
Deployment flexibility: dual-platform GPU + NPU support, already validated at scale on domestic clusters. Released simultaneously on HuggingFace, GitHub, and ModelScope.
Deeper Analysis
The real significance here is not "yet another open-source large 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: a 1.6T MoE trained on a 50K-chip domestic cluster
The LongCat team's domestic compute exploration started from a thousand chips in 2023, spending three years on operator adaptation, communication optimization, and fault-tolerant scheduling. This is an "industry first" — not "also used domestic chips," but "the first domestic cluster capable of running the full trillion-parameter pipeline." This means Chinese AI vendors now have concrete evidence of not being chokehold-dependent at the hardware layer, and the next-generation models from DeepSeek, Alibaba, and ByteDance could follow the same path.
2. Model sovereignty: 1M context + 60+ benchmark SOTA is beyond "good enough"
SWE-bench Pro 59.5 surpassing GPT-5.5 (58.6) is a representative signal that a Chinese large model has, for the first time, stably beaten an OpenAI flagship on an internationally neutral AI coding benchmark. Terminal-Bench 2.1 70.8, FORTE 73.2, BrowseComp 79.9 are SOTA results across multiple verticals. This is not "good enough in Chinese" — it is SOTA on international benchmarks.
3. Ecosystem sovereignty: native Claude Code / OpenClaw / Hermes integration
The most critical step: LongCat-2.0 was protocol-level integrated with mainstream AI coding toolchains at launch. This means Chinese large models are no longer "alternatives to foreign toolchains" but are "actively made compatible by top foreign toolchains." The "three-way binding" of Cursor, Grok, and LongCat was a first case on 06-30; by 07-05 it had become a pattern.
A three-layer engineering reality
The word "sovereignty" has been overused. LongCat-2.0 offers a clear, quantifiable definition:
> Sovereignty = the closed triangle of compute (domestic 50K chips) + model (1.6T MoE with 1M context) + ecosystem (native Claude Code/Agent protocol integration) — none of the three can be missing.
Why It Matters
Risks and Open Questions
Conclusion
Meituan's fully open-sourced LongCat-2.0 is not "just another large model." It is the most convincing engineering proof in the 2026 Chinese AI sovereignty narrative — a closed triangle of compute, model, and ecosystem, with the MIT license ensuring this combination serves not just Meituan but the entire Chinese AI industry.
Impact on the AI coding industry: the SWE-bench Pro first tier has gained a Chinese name — and it's open source. Toolchains like Cursor and Claude Code now have a SOTA option that does not depend on US hardware.
As for whether Chinese large models can stand on their own: the answer is neither "can" nor "cannot," but "already in progress, with a landmark engineering milestone delivered."
---
Sources: