Meituan Open-Sources LongCat-2.0 — 1.6 Trillion Parameter Model Trained on Chinese Chips Goes MIT License

What LongCat-2.0 Is — and Why It Matters for AI Video

Meituan — yes, the Chinese food delivery and services giant — has released LongCat-2.0 under the MIT license. The model has 1.6 trillion total parameters (around 49 billion active per inference), trained entirely on domestic Chinese AI chips. It currently sits as the top-ranked model on OpenRouter’s leaderboard for overall performance.

For video creators, the significance is not about the model itself — it is about what the model represents. LongCat-2.0 proves that competitive near-frontier capabilities can be achieved without access to NVIDIA’s GPU ecosystem. The training was done on chips from Chinese manufacturers like Huawei (Ascend series) and others in the domestic supply chain. If the US were to restrict NVIDIA chip exports further — which has been a recurring policy discussion — the existence of proven alternatives like LongCat-2.0 means AI training and inference infrastructure can continue operating outside the NVIDIA ecosystem.

How LongCat-2.0 Performs

The 49 billion active parameters per inference puts LongCat-2.0 in the same weight class as DeepSeek-V4 and similar MoE architectures. The 1.6 trillion total parameters mean the model has a vast knowledge store available through its expert routing, but only the relevant experts are activated for any given query — which keeps inference costs manageable.

On OpenRouter, users consistently rate LongCat-2.0 at the top of the leaderboard across categories including general reasoning, coding, and creative writing. The model competes with frontier offerings from Anthropic and OpenAI on many benchmarks, though specific video generation benchmarks have not been released.

What MIT License Means for AI Video Tools

LongCat-2.0 being released under the MIT license is arguably more important than the model’s benchmark scores. MIT is one of the most permissive open-source licenses — it allows anyone to use, modify, distribute, and sublicense the software with no restrictions beyond preserving the copyright notice.

For AI video tool developers and creators who rely on AI tools, this means:

• Self-hosting is legal and free. Any developer can download the model weights, set up serving infrastructure, and run LongCat-2.0 without paying licensing fees or complying with usage restrictions. This is in contrast to many near-frontier models that have commercial use restrictions, rate limits, or requirement to share derivative works.
• Fine-tuning for video-specific tasks is unrestricted. If LongCat-2.0’s underlying architecture supports video understanding or generation — or if a derivative fine-tune can add those capabilities — the MIT license imposes no barriers. The same cannot be said for many proprietary models.
• No data-sharing requirements. Unlike some open-source licenses (AGPL, for example), MIT does not require sharing modifications. Companies building proprietary video AI tools on top of LongCat-2.0 can keep their modifications private.

The Geopolitical Context: Chip Independence and AI Resilience

LongCat-2.0’s release comes at a time when the US and China are engaged in ongoing semiconductor export control negotiations. The US has restricted export of advanced NVIDIA GPUs to China. China has responded by accelerating domestic chip production and training AI models on those chips.

This creates a strategic reality for anyone building or using AI video tools: the global AI infrastructure is bifurcating. Models trained on Western chips (NVIDIA, AMD) dominate one ecosystem. Models trained on Chinese chips (Huawei Ascend, Cambricon, etc.) are emerging as a parallel track. Over time, these ecosystems will diverge in capabilities, pricing, and availability.

For video creators and businesses, the practical implication is redundancy. If you build your AI video production workflow exclusively around NVIDIA-dependent models and services, you are exposed to chip supply disruptions, export policy changes, and pricing volatility in a single ecosystem. Models like LongCat-2.0, released under permissive licenses and trained on alternative hardware, provide a hedge — they can be deployed on non-NVIDIA infrastructure and will continue to be available regardless of chip trade policy changes.

This is not a theoretical concern. The US Commerce Department has adjusted semiconductor export rules multiple times over the past three years, each time changing which chips can be exported to which customers. The trend is toward tighter restrictions, not looser. Having capable open-source models trained on non-US hardware available under permissive licenses means that AI video capabilities are not subject to the same export control risks as the hardware.

How LongCat-2.0 Fits into the AI Video Production Stack

LongCat-2.0 is primarily a language model, not a video generation model. But language models are increasingly integral to video production workflows in several ways that make LongCat-2.0 relevant to AI video creators:

Script and caption generation: LongCat-2.0’s top-ranked creative writing performance makes it a strong candidate for generating video scripts, captions, and voiceover text. Running a local or API-based instance for narrative generation avoids the per-character spending limits and content policies of using Claude or GPT for the same task.

Video metadata and search optimization: Generating SEO-optimized titles, descriptions, tags, and transcripts at scale requires a capable language model. LongCat-2.0’s performance on these tasks is comparable to frontier models at potentially lower inference cost.

Prompt engineering for video models: High-quality prompts for video generation models (like Sora, Veo, or Gen-3) require detailed scene descriptions. LongCat-2.0 can generate, refine, and iterate on these prompts more effectively than smaller models.

Code generation for video tools: For developers building custom video processing pipelines, editing tools, or automation scripts, LongCat-2.0’s coding performance on OpenRouter is competitive with dedicated code models.

Cost arbitrage: MIT-licensed models self-hosted on non-NVIDIA hardware (even modest CPU-based inference for non-real-time tasks) can be dramatically cheaper than per-token API calls to commercial model providers, especially for batch processing of large volumes of video metadata, scripts, and prompts.

What to Watch Next

LongCat-2.0 is the highest-profile example of a trend that is accelerating: capable open-source models trained on non-Western hardware. Expect to see more entries in this category over the next 12-18 months as Chinese chip manufacturers improve their training hardware and more Chinese companies follow Meituan’s lead in open-sourcing their models.

For AI video creators, the key question is how quickly these models develop multimodal capabilities. LongCat-2.0 is text-only today. A future LongCat-3.0 with native video understanding or generation would be a genuinely disruptive release — a near-frontier multimodal model, MIT-licensed, trained on chips that are not subject to US export controls.

The infrastructure for running these models is also developing rapidly. Chinese cloud providers are building inference-optimized services around domestic chips, making it easier to deploy models like LongCat-2.0 without managing hardware. And permissive licensing means third-party providers globally can offer LongCat-2.0 as a service — OpenRouter already does, and more are likely to follow.

The OpenRouter Ecosystem: What #1 Ranking Actually Means

OpenRouter is the most transparent model comparison platform in the AI industry. Unlike closed benchmarks or vendor-reported scores, OpenRouter aggregates real user feedback across thousands of diverse queries — coding problems, creative writing tasks, reasoning puzzles, factual questions, and domain-specific challenges. Users rate responses on a 5-point scale, and the leaderboard reflects aggregate satisfaction across all categories.

LongCat-2.0’s #1 ranking on this platform is significant because it represents real-world utility rather than benchmark optimization. Many models are tuned specifically to perform well on academic benchmarks like MMLU, GSM8K, or HumanEval — and then underperform in actual usage because the benchmarks don’t capture the full range of real user needs. OpenRouter’s ranking is a composite of actual user satisfaction across many task types.

For AI video creators, this means LongCat-2.0 is likely a reliable tool for the surrounding tasks that video production requires: script generation, structured prompt creation, metadata optimization, and code generation for video processing pipelines. The model’s consistent performance across task types reduces the need to switch between different models for different parts of the workflow — one model can handle narrative, technical, and creative tasks with equal competence.

The pricing dynamics are also favorable. On OpenRouter, LongCat-2.0 is priced competitively against equivalent-tier models. For bulk processing — generating hundreds of video scripts, SEO-optimized descriptions, or prompt variants — the per-token cost advantage compounds significantly. A creator processing 100 product descriptions for e-commerce videos, each requiring a 500-token generation, saves roughly 40-60% compared to using the same volume through Claude or GPT APIs, depending on the specific pricing tier.

Practical Implications for Video Creators Outside the US

For video creators, AI tool developers, and content producers outside the United States, LongCat-2.0’s release has specific practical implications that go beyond the technology itself.

API availability independent of US policy: Meituan operates out of China. The company has no obligation to comply with US export controls or sanctions regimes. A creator or business that relies on models served by Chinese companies or through Chinese cloud providers has an API that will not be disrupted by US policy changes. This is valuable redundancy for international operations that depend on consistent AI service availability.

Pricing independence: The AI model market is currently dominated by US companies that price in dollars and set terms largely based on US market conditions. A competitive Chinese-origin model under permissive license creates pricing pressure. When a near-frontier model is available at competitive rates under MIT license, it constrains how much US providers can raise prices without losing market share — which benefits every AI tool user regardless of geography.

Self-hosting capability: MIT license + training on domestic chips means LongCat-2.0 can be self-hosted anywhere in the world without depending on cloud providers that may have regional restrictions or data residency concerns. For video creators in markets with strict data sovereignty laws (GDPR in Europe, PIPL in China, PDPL in Saudi Arabia), self-hosting a model like LongCat-2.0 provides AI capabilities that comply with local data regulations.

Community ecosystem: Open-source models develop communities of developers, fine-tuners, and tool builders who extend the base model’s capabilities. LongCat-2.0’s MIT license makes it eligible for the broadest possible ecosystem growth. Expect third-party tools, fine-tunes, and integrations to emerge faster for MIT-licensed models than for restricted-license alternatives.

The trajectory is clear. Open-source AI models are becoming more capable, more accessible, and more independent of specific hardware ecosystems. LongCat-2.0 is the current state of the art in this category, but it will not hold that position alone. The release strategy — open weights, permissive license, domestic chips — will be replicated by other Chinese companies and, increasingly, by non-US AI developers globally. The result is a more resilient, more competitive, and more accessible AI ecosystem for everyone who uses AI tools for creative and productive work.

FAQ

About Barak Schurr

Barak Schurr is a serial entrepreneur who built and exited a major ticketing platform before turning his attention to AI and content. He runs three AI-focused research sites and writes about the tools, workflows, and strategies that let solo creators build real channels — without a crew, a studio, or a Hollywood budget.
Email: [email protected]

More from the Barak Schurr Network

Top Notch AI Tools

Practical AI tools, guides & resources for digital entrepreneurs

AI Table Review

AI strategies and tools for independent restaurant owners and operators