LongChat 2.0: The 1.6 Trillion Parameter Model Trained Without Nvidia GPUs
Meituan's LongChat 2.0 is a 1.6T parameter open-weight model trained on custom AI chips—no Nvidia GPUs required. Here's how they did it and why it matters.
What Meituan Just Shipped — And Why It Matters
A 1.6 trillion parameter open-weight model trained entirely without Nvidia GPUs. That sentence would have seemed improbable two years ago. Today, it’s the headline from Meituan, China’s largest food delivery and local services platform, with the release of LongChat 2.0.
The model is significant for reasons that go beyond its raw scale. LongChat 2.0 demonstrates that frontier-class LLMs can be developed outside the Nvidia ecosystem — a practical proof point at a moment when US export controls have cut off Chinese AI labs from H100 and A100 chips. The fact that it’s open-weight makes it even more noteworthy: the weights are publicly available, which means anyone can examine, fine-tune, or deploy the model.
This article breaks down what LongChat 2.0 actually is, how Meituan built it on alternative hardware, what the benchmarks look like, and what it signals for the broader enterprise AI landscape.
LongChat 2.0: The Basics
Meituan is not primarily an AI company. It’s a tech conglomerate — best known for food delivery, hotel bookings, and local commerce — with over 700 million annual transacting users. But it has been quietly building out AI infrastructure, and LongChat 2.0 is the public-facing result of that work.
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
LongChat 2.0 is a Mixture of Experts (MoE) model with 1.6 trillion total parameters. That number is the aggregate across all expert sub-networks. At inference time, only a fraction of those parameters activate for any given input — the routing mechanism selects which experts to engage. This architecture is the same approach used by DeepSeek-V3 and Mixtral, and it’s what makes a 1.6T parameter model practically trainable and deployable without the full compute cost you’d associate with a 1.6T dense model.
The model supports long-context inputs — hence the name — with a context window that extends well beyond what was standard even a year ago. This makes it suited for tasks like document analysis, multi-turn dialogue over long sessions, and retrieval-augmented generation pipelines where large chunks of text need to fit in context.
Meituan released LongChat 2.0 as an open-weight model, meaning the model weights are publicly available. This is distinct from open-source (which would include training code, data, and methodology), but it’s still a meaningful degree of openness — developers can download, run, and fine-tune the model without going through a proprietary API.
The Architecture: Why MoE Makes This Scale Possible
To understand how a company like Meituan gets to 1.6 trillion parameters without a datacentre full of H100s, you need to understand what Mixture of Experts actually does.
How MoE Works
In a standard dense transformer, every parameter activates for every token. Scale to 1.6 trillion parameters and the compute cost becomes prohibitive at both training and inference.
MoE breaks the feed-forward layers of the transformer into multiple parallel “expert” networks. A learned routing function decides which experts handle each token — typically 2–8 out of a larger pool. The rest stay idle. The result: a model with a very large parameter count but a much smaller “active parameter count” that actually does work on any given forward pass.
DeepSeek-V3, for reference, has 671 billion total parameters but only 37 billion active parameters per forward pass. LongChat 2.0 follows the same logic at a larger total scale.
What This Means for Training
MoE models are harder to train than dense models. The routing mechanism needs to balance load across experts — if all tokens route to the same few experts, you’ve effectively trained a much smaller model. Getting this right requires careful auxiliary loss functions and expert capacity tuning.
But the payoff is real: you can train a model with far more total capacity than compute budget would otherwise allow. MoE also creates natural opportunities for parallelization, which matters a lot when your training hardware isn’t the fastest chips on the market.
Training Without Nvidia: What Hardware Did Meituan Use?
This is the part that’s generating the most attention. US export controls, tightened progressively since 2022, have blocked Chinese companies from purchasing Nvidia’s most advanced datacenter GPUs — including the H100, A100, and their successors. Companies that had already stockpiled chips before the restrictions are working through those reserves. Everyone else is building alternatives.
The Huawei Ascend Ecosystem
The most widely used alternative in China is Huawei’s Ascend line of AI accelerators, particularly the Ascend 910B and 910C. These chips were designed specifically to fill the gap left by restricted Nvidia hardware. Huawei has invested heavily in both the hardware itself and the software stack — its CANN (Compute Architecture for Neural Networks) framework is the primary tool for model training and inference on Ascend hardware.
Meituan’s LongChat 2.0 training run used Ascend-based infrastructure. This is consistent with what other major Chinese AI labs have reported: Baidu, ByteDance, and others have all shifted significant training workloads to Ascend hardware over the past 18 months.
The Engineering Challenges
Training at this scale on non-Nvidia hardware is not a drop-in substitution. The software ecosystem around Nvidia’s CUDA is mature, well-documented, and has benefited from years of optimization by thousands of researchers. Huawei’s CANN is capable, but the tooling, debugging infrastructure, and community knowledge base are not yet at the same level.
Teams building on Ascend typically need to:
- Port model code from PyTorch/CUDA to CANN or use compatibility layers
- Retune memory management and kernel fusion optimizations for the Ascend architecture
- Develop custom communication patterns for large-scale distributed training
- Handle reliability issues more manually, since the ecosystem for monitoring large training runs is less mature
The fact that Meituan completed a 1.6T parameter training run on this stack is a genuine engineering accomplishment — separate from any questions about benchmark performance.
What This Signals for the Chip Market
The broader implication is straightforward: the gap between Nvidia and its alternatives is narrowing, at least for inference quality if not yet for raw training efficiency. Export controls have accelerated domestic chip development in ways that are already showing up in production systems. LongChat 2.0 is one data point. DeepSeek’s runs on domestic hardware are another.
This doesn’t mean Nvidia is under immediate competitive threat globally. But it does mean that the assumption “frontier AI requires Nvidia hardware” is becoming less absolute.
How LongChat 2.0 Performs
Benchmark performance for models of this size is always context-dependent. The numbers matter less than the gap to comparable models and how they hold up on domain-specific tasks.
Benchmark Results
Meituan’s published benchmarks show LongChat 2.0 performing competitively on:
- MMLU (general knowledge and reasoning): scores in the range of leading open-weight models
- HumanEval (code generation): strong results, particularly relevant given Meituan’s engineering-heavy use case
- Long-context tasks: performance on tasks requiring comprehension of 32K+ token inputs is a reported strength, aligned with the model’s design goals
- Chinese-language benchmarks: top-tier performance, which is expected given Meituan’s training data composition
Direct comparisons against GPT-4o or Claude 3.5 Sonnet show LongChat 2.0 within striking distance on many tasks, though not uniformly ahead. This is a reasonable outcome for a model of this class.
The Long-Context Claim
The “LongChat” branding is intentional. The model is specifically optimized for tasks where context length is the binding constraint. For enterprise use cases — legal document review, long customer service conversation histories, technical documentation analysis — this is a meaningful differentiator.
Many open-weight models nominally support long contexts but degrade significantly in accuracy as context fills up. Whether LongChat 2.0 holds up consistently across its full context window is something third-party evaluations will need to verify over time.
Why an Open-Weight Model From Meituan Is Interesting
Meituan releasing an open-weight model is not an obvious business move. Unlike a research lab, they have a concrete product portfolio to protect. So why open-weight?
The Research Credibility Play
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
Publishing model weights invites scrutiny and benchmarking from the broader community. If LongChat 2.0 holds up, that’s a signal to the research community that Meituan is a serious AI organization — which helps with recruiting and partnerships.
The Ecosystem Play
Open-weight models build ecosystems. Developers who fine-tune and build on LongChat 2.0 create a community that’s invested in Meituan’s AI stack. This is the same logic Meta used with Llama.
The Regulatory Context
There’s also a geopolitical dimension. Releasing open-weight models demonstrates AI development capability in a way that’s visible and credible to Chinese regulators and the international research community. In the current environment, that has real value.
Enterprise AI Implications
For enterprise AI teams, LongChat 2.0 is worth tracking for a few specific reasons.
Supply Chain Diversification
If you’re building AI systems at scale, your current stack probably has Nvidia at the center. LongChat 2.0 and the infrastructure behind it are evidence that viable alternatives exist. This matters for risk management — dependency on a single hardware vendor for mission-critical AI infrastructure is a real concern.
Open-Weight Models for Sensitive Workloads
Open-weight models can be deployed on-premises, which is the only acceptable option for many regulated industries. A 1.6T parameter model with strong long-context performance is potentially useful for legal, financial, and healthcare applications where data cannot leave controlled infrastructure.
Fine-Tuning at Scale
The open-weight release means enterprise teams can fine-tune LongChat 2.0 on proprietary data. For organizations with large Chinese-language corpora or specific domain knowledge, this is a practical starting point for customized models that outperform general-purpose alternatives.
Where MindStudio Fits Into This Picture
The release of powerful open-weight models like LongChat 2.0 is only useful if teams can actually integrate them into real workflows. That’s consistently the harder problem: not which model to use, but how to build reliable, production-grade AI applications on top of them.
MindStudio is a no-code platform for building and deploying AI agents and automated workflows. It gives teams access to 200+ AI models out of the box — including open-weight models and proprietary APIs — without needing to manage API keys, infrastructure, or model-specific integration code.
The practical value here is flexibility. As the open-weight model landscape evolves — and LongChat 2.0 is a good example of that evolution accelerating — teams shouldn’t have to rebuild their agent infrastructure every time a new best-in-class model appears. MindStudio’s model-agnostic architecture means you can swap the underlying model in a workflow without rebuilding the logic around it.
For long-context use cases specifically — document processing, multi-turn dialogue agents, retrieval-augmented pipelines — MindStudio’s visual workflow builder lets you chain model calls, manage context windows, and connect to data sources through 1,000+ pre-built integrations, without writing infrastructure code. The average build takes 15 minutes to an hour.
If you’re evaluating models like LongChat 2.0 for enterprise use cases, you can start building on MindStudio for free at mindstudio.ai.
For more on how enterprise teams are approaching open-weight model deployment, see MindStudio’s guide to building AI agents for business workflows.
Frequently Asked Questions
What is LongChat 2.0?
LongChat 2.0 is a large language model developed by Meituan, the Chinese technology company. It has 1.6 trillion total parameters arranged in a Mixture of Experts (MoE) architecture, meaning only a subset of those parameters activate per inference. It is released as an open-weight model, allowing developers to download and fine-tune the weights. It is optimized for long-context tasks, supporting input sequences significantly longer than many comparable models.
How was LongChat 2.0 trained without Nvidia GPUs?
LongChat 2.0 was trained on Huawei Ascend AI accelerators, which are the primary domestically-produced alternative to Nvidia’s datacenter GPUs in China. US export controls prevent Chinese companies from purchasing Nvidia’s most advanced chips (H100, A100, and their successors), which has driven investment in Huawei’s Ascend ecosystem. Training on Ascend requires custom software work to port standard ML frameworks from Nvidia’s CUDA stack, but the hardware is now capable enough to run training runs at this scale.
What does “open-weight” mean, and how is it different from open-source?
An open-weight model is one where the trained model weights are publicly released, allowing anyone to run and fine-tune the model. Open-source goes further — it includes training code, datasets, and full methodology documentation. LongChat 2.0 is open-weight: you can download and use the model, but Meituan hasn’t released all the details of the training process. This is the same model release strategy used by Meta with Llama.
How does the Mixture of Experts architecture work?
In a Mixture of Experts model, the feed-forward layers of the transformer are split into multiple parallel “expert” networks. A routing function assigns each token to a small subset of experts — typically 2 to 8 out of a larger pool. The other experts remain inactive. This lets a model have a very large total parameter count while keeping the compute cost per forward pass manageable. DeepSeek-V3, Mixtral, and Grok-1 all use this same approach.
Is LongChat 2.0 competitive with models like GPT-4o or Claude?
On published benchmarks, LongChat 2.0 performs competitively with leading frontier models on several tasks, particularly in Chinese-language evaluation and long-context comprehension. It doesn’t uniformly outperform GPT-4o or Claude 3.5 Sonnet across all tasks, but it’s in the same performance tier — which is notable given the hardware constraints under which it was trained. Independent third-party evaluations will provide a clearer picture over time.
Why does it matter that a non-AI company built this model?
Meituan is primarily a food delivery and local commerce platform, not an AI research lab. Its ability to train a 1.6T parameter model on non-Nvidia hardware suggests that the capability to build frontier-class models is diffusing beyond specialist AI organizations. This is relevant for enterprise AI strategy: it means more competition, more open-weight options, and continued pressure on model providers to differentiate on more than raw scale.
Key Takeaways
- LongChat 2.0 is a 1.6 trillion parameter MoE model from Meituan, released as open-weight, optimized for long-context tasks.
- It was trained on Huawei Ascend chips, not Nvidia GPUs — a meaningful demonstration that frontier AI training is possible outside the Nvidia ecosystem.
- MoE architecture makes the scale practical: only a fraction of parameters activate per inference, keeping compute costs manageable.
- The open-weight release enables enterprise fine-tuning on proprietary data, which is valuable for regulated industries and domain-specific applications.
- The broader signal: US export controls are accelerating alternatives to Nvidia hardware in ways that are now showing up in production models at scale.
For teams building AI applications, platforms like MindStudio make it practical to work with multiple models — including emerging open-weight options — without rebuilding infrastructure every time the model landscape shifts. You can explore AI agent building on MindStudio to see how it handles model flexibility in practice.
