Meituan, the Chinese technology conglomerate best known internationally for its food-delivery empire, has released and open-sourced LongCat-2.0, a 1.6 trillion-parameter artificial intelligence model purpose-built for agentic software engineering. The release, announced on June 30, 2026, is notable not only for its scale but for how it was trained: entirely on domestically produced Chinese chips, with no Nvidia or AMD accelerators in the pipeline.
The move is being read as the strongest signal yet that US export controls on advanced AI silicon are failing to prevent Chinese firms from producing frontier-class models. Multiple outlets, including the South China Morning Post, Reuters and VentureBeat, reported the launch, with Crypto Briefing describing LongCat-2.0 as one of the largest open-source coding models ever built. For more context on this story, see our ongoing breaking AI news.
A Mixture-of-Experts model built for agents
LongCat-2.0 uses a Mixture-of-Experts (MoE) architecture, meaning it does not activate all 1.6 trillion parameters on every request. Instead, it dynamically routes each token through roughly 33 billion to 56 billion active parameters, keeping it far more efficient than a comparably sized dense model while retaining a sprawling knowledge base to draw from.
That efficiency matters for cost. MoE designs have become the dominant architecture for large-scale models precisely because they decouple total capacity from per-query compute, allowing labs to ship very large models without charging customers for every parameter on every call.
The model supports a context window of 1 million tokens, large enough to ingest an entire substantial codebase and reason across it in a single pass — a capability increasingly demanded by autonomous coding agents that must hold whole repositories in mind rather than work file by file.
Benchmarks and real-world usage
On benchmarks designed to measure real-world software engineering, LongCat-2.0 posted a score of 59.5 on SWE-bench Pro and 70.8 on Terminal-Bench, according to Crypto Briefing. These evaluation suites test how well models handle the messy, multi-step challenges of genuine development work — fixing bugs, writing features, navigating existing code — rather than simple one-shot programming puzzles.
VentureBeat separately noted that the model had been "leading OpenRouter," the popular multi-model API gateway, in agentic coding rankings ahead of its formal open-source release. OpenRouter aggregates more than 400 models from over 60 providers and its usage data is widely treated as an industry barometer of which models developers actually choose to run.
Trained on 50,000 domestic chips
The detail drawing the most attention is the hardware. Meituan said it trained and ran LongCat-2.0 on a 50,000-card domestic compute cluster built entirely from Chinese-manufactured accelerators. According to Crypto Briefing, that means no Nvidia A100s or H100s and no AMD MI300X units were used at any stage of training — a clean break from the hardware that has powered most large-model development worldwide.
Meituan claims the 50,000-card cluster represents a first at this scale for domestic Chinese hardware. The training run was completed despite US export controls that have systematically restricted China's access to cutting-edge AI chips since late 2022. The implication, analysts say, is that the controls' leverage is eroding as Chinese chipmakers and their customers close the performance gap.
Weights available on Hugging Face
In keeping with the open-source framing, Meituan has published the model weights and supporting resources on Hugging Face under the "meituan-longcat" organization. That lowers the barrier for developers and researchers who want to fine-tune the model for enterprise codebases, security auditing or automated debugging but could never afford to train something comparable from scratch.
The open release also feeds directly into the agentic-coding ecosystem, where open-weights models can be self-hosted, audited and modified without the per-token costs and data-handling concerns that come with proprietary APIs — a significant consideration for enterprises wary of sending proprietary code to a third-party cloud.
A rapid model lineage
LongCat-2.0 did not appear overnight. The lineage traces back to LongCat-Flash, a 560 billion-parameter model released in September 2025, followed by LongCat-Next, a multimodal variant, in March 2026. The new release nearly triples the parameter count of its predecessor in under a year, reflecting the breakneck pace at which Chinese labs are now iterating on foundation models.
The launch also comes amid a broader reshuffling of the global model leaderboard, in which Chinese systems have surged in real-world developer usage and now top token-consumption rankings on aggregator platforms.
What it means for the chip-control debate
For policymakers in Washington, LongCat-2.0 is an uncomfortable data point. The central premise of the export-control strategy has been that denying access to leading-edge chips would slow China's AI progress. Every model trained successfully on domestic hardware weakens that premise, suggesting the restrictions may be accelerating the very self-sufficiency they were meant to prevent.
Meituan itself has struck a measured public tone. Founder and chief executive Wang Xing recently reiterated the company's commitment to AI investment while stressing it would not "blindly invest beyond its financial capacity," according to reports carried by Futu.
For now, the takeaway for developers is concrete: a near-frontier, million-token coding model is freely downloadable, and it was built without a single restricted American chip. Whether US labs respond with cheaper open weights, tighter controls, or a renewed push for raw performance, the competitive landscape just shifted again.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
