China's AI Cambrian Explosion: Tencent and Meituan Unleash Competing Open-Weight Giants

# China's AI Cambrian Explosion: Tencent and Meituan Unleash Competing Open-Weight Giants By Wei Lian | Neuron China Desk | July 9, 2026

The first week of July 2026 has marked a pivotal moment in China's artificial intelligence race, transforming it from a pursuit of Western benchmarks into a fierce, multi-fronted domestic rivalry. Within days of each other, two of the nation's most formidable tech titans, Tencent and Meituan, unleashed major new open-weight foundation models. Tencent formally released its highly optimized **Hunyuan Hy3**↗, while Meituan stunned the industry with **LongCat-2.0**↗, a colossal 1.6 trillion-parameter model trained entirely on domestic hardware.

These are not merely iterative updates. The near-simultaneous launches of Hy3 and LongCat-2.0 represent the crystallization of two distinct and sophisticated strategies for AI dominance. Tencent is flexing its muscles as an integrated cloud and software giant, offering a cost-effective, enterprise-ready model designed to lock developers into its vast ecosystem. Meituan, a sprawling e-commerce and local services platform, has made a powerful statement on technological sovereignty, proving that near-frontier capabilities can be achieved without reliance on foreign compute infrastructure. This Neuron China Desk report unpacks the technical specifications, performance claims, and strategic implications of these landmark releases, which signal a new, more intense phase of competition within China's walled garden.

The New Contenders: Cloud Ecosystem vs. Sovereign Compute

The duel between Tencent's Hy3 and Meituan's LongCat-2.0 is a microcosm of the broader strategic schisms emerging in China's AI landscape. Tencent, the operator of WeChat and a dominant force in cloud computing and gaming, views AI as the connective tissue for its sprawling empire. Its strategy is one of deep integration, where Hunyuan serves as the intelligent engine powering everything from its enterprise collaboration tool, WorkBuddy, to its consumer-facing AI assistant, Yuanbao. Releasing a highly capable and aggressively priced model like Hy3 is a direct play to attract developers to the Tencent Cloud platform, commoditizing the model layer to sell higher-margin cloud services and solidifying its role as a fundamental utility provider for the digital economy.

Meituan's approach is born from a different necessity and ambition. As a platform whose core business involves the complex logistics of food delivery, travel booking, and local commerce, Meituan's interest in AI is existential. However, its release of LongCat-2.0 transcends immediate business needs. By explicitly training the model on a massive cluster of over 50,000 domestic Huawei Ascend AI chips, Meituan is not just building a model; it is building a proof case for China's entire sovereign AI stack. This move directly addresses the nation's strategic vulnerability to foreign semiconductor restrictions and demonstrates to the world — and to Beijing — that a parallel path to AI leadership is viable. This makes LongCat-2.0 as much a geopolitical statement as it is a technological achievement.

The near-simultaneous launches of Hunyuan Hy3 and LongCat-2.0 are not just product releases — they are strategic declarations. One says: "We can out-price the West." The other says: "We don't need the West's chips."

Deep Dive: Tencent's Hunyuan Hy3

Tencent's official launch of Hunyuan Hy3↗ on July 6, 2026, marks the model's graduation from a preview version released in April to a fully-fledged, production-ready offering. The most significant change is the adoption of the permissive Apache 2.0 license. This move removes the geographic and field-of-use restrictions of the preview, signaling Tencent's intent for Hy3 to become a global, freely-usable open-weight standard.

Architecturally, Hy3 is a formidable example of modern efficiency. It is a Mixture-of-Experts (MoE) model with a total of 295 billion parameters, of which a lean 21 billion are activated per token. This dense-MoE hybrid structure, composed of 192 routed experts and one shared expert per layer, is engineered for cost-effective inference without sacrificing significant capability. It supports a 256K context window and incorporates a 3.8 billion-parameter multi-token prediction (MTP) layer to accelerate speculative decoding and reduce latency.

Hy3 Key Specifications at a Glance

• Architecture: Mixture-of-Experts (MoE) with 295B total parameters, 21B active per token — delivering frontier-class reasoning at a fraction of the inference cost of dense models.
• Context window: 256,000 tokens, sufficient for processing entire codebases, lengthy legal documents, or extended multi-turn agentic sessions without truncation.
• License: Apache 2.0 — fully permissive for commercial use globally, with no geographic restrictions, a deliberate upgrade from the preview's more limited terms.
• Pricing on Tencent Cloud: approximately ¥1 ($0.14 USD) per million input tokens and ¥4 ($0.56 USD) per million output tokens; cached input drops to just ¥0.25, making context-heavy applications extremely cost-effective.
• Benchmark positioning: 90% task-resolution rate on Tencent's internal "WorkBuddy" enterprise assistant; competitive with general reasoning flagships, though Tencent acknowledges specialized coding models like GLM-5.2 currently lead on pure code benchmarks.

In terms of performance, Tencent has been transparent about Hy3's strengths and weaknesses. The company claims the model achieves an impressive 90% task-resolution rate on its internal "WorkBuddy" enterprise assistant, showcasing its proficiency in complex, multi-turn reasoning and agentic workflows. While competitive with other flagship models in general reasoning, Tencent acknowledges that specialized models like Zhipu's GLM-5.2 currently hold an edge in pure coding-heavy benchmarks. This positioning suggests Hy3 is a versatile "all-rounder" optimized for the broad spectrum of enterprise tasks that dominate cloud workloads.

This strategy is clearly reflected in its aggressive pricing. On Tencent Cloud, Hy3 is offered at approximately ¥1 ($0.14 USD) per million input tokens and ¥4 ($0.56 USD) per million output tokens. For context-heavy applications, the price for cached input drops to just ¥0.25. This pricing structure is designed to aggressively undercut both international competitors and domestic rivals, making it an economically irresistible option for developers already within or considering entering the Tencent Cloud ecosystem.

Meituan's Marvel: The 1.6 Trillion-Parameter LongCat-2.0

Meituan's announcement of LongCat-2.0↗ on June 30, followed by the release of its open weights under an MIT license on July 4, sent shockwaves through the AI community. The model's specifications are staggering: a 1.6 trillion-parameter MoE architecture that activates an average of 48 billion parameters per token and boasts a native 1 million-token context window.

The most profound aspect of LongCat-2.0 is its provenance. Meituan has confirmed the model was trained and is served entirely on domestic hardware, making it arguably the most powerful AI model in the world not reliant on NVIDIA GPUs. As VentureBeat reported↗, this achievement is a landmark success for China's efforts to build a self-sufficient AI supply chain.

LongCat-2.0 Key Specifications at a Glance

• Architecture: MoE with 1.6 trillion total parameters and ~48B active per token — the largest open-weight model trained exclusively on domestic Chinese chips, surpassing even GLM-5.2's 744B in raw scale.
• Context window: 1 million tokens natively, enabled by the proprietary LongCat Sparse Attention (LSA) mechanism — an evolution of DeepSeek's sparse attention that achieves linear-complexity scaling.
• License: MIT — the most permissive open-source license available, allowing unrestricted commercial use, modification, and redistribution without attribution requirements.
• Benchmark: 59.5 on SWE-bench Pro, narrowly surpassing GPT-5.5's reported score of 58.6, establishing near-frontier status in software engineering tasks.
• Pricing: Standard rate of $0.75 per million input tokens and $2.95 per million output tokens; promotional rates of $0.30 and $1.20 respectively; cached context reads provided free of charge.
• Pre-release stealth: Operated anonymously on OpenRouter as "Owl Alpha," ranking among the top three models globally by call volume on developer tools like Hermes Agent before its identity was revealed.

LongCat-2.0 is positioned as a specialist in agentic coding. Meituan's own benchmarks show it achieving a score of 59.5 on SWE-bench Pro, narrowly surpassing the reported score for GPT-5.5 (58.6) and demonstrating its near-frontier status in software engineering tasks. The model's technical foundation is equally impressive, featuring proprietary innovations like LongCat Sparse Attention (LSA), an evolution of DeepSeek's sparse attention mechanism that enables linear-complexity scaling for its massive context window. Prior to its official reveal, the model operated anonymously on the OpenRouter platform↗ under the codename "Owl Alpha," rapidly gaining popularity and ranking among the top three models globally by call volume on developer tools like Hermes Agent and Claude Code.

Meituan's pricing strategy for LongCat-2.0 is designed to be highly competitive, with a standard rate of $0.75 per million input tokens and $2.95 per million output tokens, and an even lower promotional rate. Remarkably, Meituan announced that cached context reads would be provided free of charge, a move that drastically lowers the cost for applications involving long, persistent contexts, such as codebase analysis or multi-turn agentic conversations.

LongCat-2.0's anonymous debut as "Owl Alpha" on OpenRouter — where it ranked in the global top three by call volume before anyone knew its origin — is perhaps the most telling detail of all. The model earned its reputation on merit alone, before the Meituan brand was attached.

Side-by-Side: Hy3 vs. LongCat-2.0

| Feature | Tencent Hunyuan Hy3 | Meituan LongCat-2.0 | | :--- | :--- | :--- | | Official Release | July 6, 2026 | June 30, 2026 (Weights July 4) | | Total Parameters | 295 Billion (MoE) | 1.6 Trillion (MoE) | | Active Parameters | 21 Billion | ~48 Billion (Average) | | Context Window | 256,000 tokens | 1,000,000 tokens | | License | Apache 2.0 | MIT | | Positioning | General Enterprise, Reasoning Agent | Specialist Agentic Coding | | Training Hardware | Not Specified (Presumed Mixed) | Domestic Ascend Chips (Exclusively) | | Key Benchmark | 90% task resolution on "WorkBuddy" | 59.5 on SWE-bench Pro | | Standard Pricing (Input) | ~$0.14 / 1M tokens | $0.75 / 1M tokens (Promo: $0.30) | | Standard Pricing (Output)| ~$0.56 / 1M tokens | $2.95 / 1M tokens (Promo: $1.20) |

Honorable Mentions and the Broader Landscape

While the Tencent and Meituan releases captured the spotlight, they were not the only significant developments. On July 7, Ant Group's robotics division, Robbyant, announced its **LingBot-Vision** and **LingBot-Depth 2.0** models↗, open-sourcing the weights for the former. This launch focuses on embodied AI, targeting the difficult problem of spatial perception for robots, a crucial but distinct area of AI development. Furthermore, industry chatter points to an imminent official launch of DeepSeek-V4 in mid-July, which is expected to introduce a dynamic peak-hour pricing mechanism to manage demand — another sign of a maturing, commercially-focused market.

The broader industry news summary from July 9↗ confirms that the pace of Chinese AI releases shows no sign of slowing. The domestic competitive pressure is now so intense that labs are releasing major models within days of each other, each trying to capture developer mindshare before the next announcement cycle begins.

Why This Matters for the West

The launches of Hunyuan Hy3 and LongCat-2.0 are a clear signal that the Chinese AI industry has entered a new era of maturity and self-confidence. For Western observers, developers, and policymakers, there are several critical takeaways.

First, the emergence of a viable, at-scale domestic compute alternative to NVIDIA hardware, as demonstrated by Meituan, fundamentally alters the geopolitical calculus of technology controls. It suggests that while such controls may slow progress, they are unlikely to halt it and may instead accelerate the development of a resilient, parallel AI ecosystem in China.

Second, the intense domestic competition is a powerful engine for innovation and price reduction. The aggressive pricing and open-licensing models offered by Tencent and others create a "buyer's market" for developers globally. This will place significant downward pressure on the pricing of Western proprietary models, forcing labs like OpenAI and Anthropic to compete not just on capability but also on cost-effectiveness, especially for the vast majority of enterprise tasks that do not require the absolute bleeding edge of performance.

Finally, the strategic divergence between Tencent and Meituan illustrates that there is no single "Chinese AI strategy." Instead, the market is a dynamic arena where different corporate and national priorities are being played out. Understanding these nuances is crucial for anyone seeking to compete with, partner with, or regulate the powerful new players emerging from China's AI crucible.

How to Access These Models Today

Both models are available now for developers who want to evaluate them:

• Hunyuan Hy3 is available via Tencent Cloud's official announcement↗ and the Tencent Cloud API. The Apache 2.0 license means weights can be downloaded and run locally without restriction.
• LongCat-2.0 weights are available on Hugging Face under the MIT license. The model can also be accessed via Meituan's API at the promotional pricing rates, or through OpenRouter where it previously ran as "Owl Alpha."
• LingBot-Vision from Ant Group's Robbyant division is open-sourced and available for embodied AI research applications.
• Developers seeking a direct comparison should note that both Hy3 and LongCat-2.0 are compatible with standard OpenAI-compatible API clients, lowering the barrier to switching or A/B testing.

The week of July 6-9, 2026 will be remembered as the moment China's AI race stopped being a story about catching up to the West and became a story about competing with itself — at a pace and scale that the rest of the world is only beginning to comprehend.