Tencent Unveils Hy4 Model Amid Cloud Push and Chip Access Questions

A Model Built for Everyday Cloud Workloads

Tencent introduced its Hy4 large language model on 4 November, positioning the system as a bridge between open-weight accessibility and performance in agent-driven tasks. The 770-billion-parameter architecture adds multimodal vision for the first time in the Hy line, enabling the model to parse and interpret images alongside text. Tencent prices Hy4 between the off-peak and peak tiers of DeepSeek V4 Pro, a model known for aggressive cost positioning, signalling that the company is targeting developers who need reliable inference at scale without committing to proprietary APIs.

The firm plans to release Hy4 Flash, a faster variant optimised for routine queries, though it has not set a public date for distributing the full Hy4 weights. That pattern mirrors the strategy of Chinese AI labs over the past year: release parameters under permissive licences, keep training data private, and funnel adoption toward affiliated cloud platforms. Tencent's cloud division has lagged Alibaba and Huawei in market share; a competitive open-weight offering gives enterprise customers a reason to provision compute through Tencent's infrastructure.

Performance Across Agent Benchmarks

In agentic coding tests, Hy4 outperformed Anthropic's Claude Opus 5.5. It also surpassed both Claude Opus 5.5 and OpenAI's GPT 6.1 Sol in agentic search and working-agent scenarios, which measure a model's ability to chain tool calls, retrieve external data, and execute multi-step workflows. Those results matter for software teams building autonomous agents that interact with APIs, databases, and version-control systems; latency and reliability in tool use often determine whether a model can replace human orchestration.

Yet on AI-assisted research tasks, Hy4 fell well behind the same closed-source systems and trailed Moonshot AI's Kimi K3.1 Pro, a 2.8-trillion-parameter open-weight model released in late October. Tencent acknowledged the gap in a statement accompanying the launch, noting that internal benchmarks covering frontier mathematics, long-horizon software development, model iteration, and visual question answering still show "a considerable distance" between Hy4 and leading proprietary offerings. The company said it will scale both pre-training compute and post-training refinement to close that distance, a roadmap that implies sustained capital expenditure on GPU clusters and data pipelines.

Independent rankings illustrate the competitive landscape. On Artificial Analysis, which recently revised its evaluation framework, Kimi K3.1 Pro ranks ninth overall and holds the top position among Chinese models. Alibaba's Qwen 3.8 Max and DeepSeek V4.1 Flash occupy the thirteenth and fourteenth slots, placing them second and third within the domestic cohort. Hy4 has not yet appeared in those tables, either because Tencent has not submitted the model for formal evaluation or because embargo periods remain in effect.

Security Constraints and Training Exclusions

Tencent removed offensive cybersecurity content from Hy4's training corpus, filtering out material related to vulnerability exploitation, reverse engineering, and penetration testing. The decision reflects growing concern among regulators and enterprise buyers that frontier models can accelerate attack workflows. Anthropic has restricted its Claude Mythos series to defensive security applications, authorised enterprise testing, and critical infrastructure protection, declining to offer public API access. While Tencent's exclusion policy is narrower, it signals awareness that open-weight releases carry reputational and liability risks if the models enable malicious use at scale.

Cybersecurity experts have documented cases in which attackers use large language models to automate reconnaissance, generate polymorphic malware, and craft phishing messages. Removing explicit offensive examples from training data does not guarantee that a model cannot be fine-tuned or prompted toward harmful outputs, but it raises the floor for adversaries who lack the resources to curate their own datasets. For Tencent, the trade-off is straightforward: a model trained on sanitised data may score lower on certain red-team benchmarks yet avoids the regulatory and public-relations fallout that follows high-profile misuse.

Leadership Reshuffle and the Hy Series

Tencent hired Yao Shunyu, a former OpenAI researcher, as chief AI scientist in late 2025, reporting directly to president Martin Lau. At that point the company was perceived to trail ByteDance, Alibaba, and Baidu in the race to ship competitive foundation models. Yao has since overseen the Hy series, launching the flagship Hy3 in April and now Hy4. The appointment underscores Tencent's willingness to import talent from US labs, even as geopolitical friction constrains collaboration on training infrastructure and export-controlled hardware.

The Hy series sits within a broader portfolio that includes embedding models, speech-recognition systems, and recommendation engines embedded in WeChat, QQ, and Tencent Video. Integrating a capable general-purpose LLM into those products could unlock new advertising formats, conversational commerce, and content moderation at scale. Whether Tencent deploys Hy4 internally or reserves it for external cloud customers remains unclear; the company's announcement focused on benchmarks rather than product roadmaps.

Chip Access and Regional Infrastructure

On the same day Tencent announced Hy4, reports emerged that Oracle had cancelled a five-year lease that would have granted Tencent access to as many as 100,000 advanced AI accelerators across data centres in Southeast Asia. Tencent declined to comment on the cancellation. If accurate, the move would constrain the firm's ability to scale inference and fine-tuning workloads outside mainland China, where US export controls limit the sale of cutting-edge chips from Nvidia and AMD.

Southeast Asia has become a focal point for Chinese cloud providers seeking to serve customers in markets with fewer regulatory obstacles. Losing access to a large cluster would force Tencent to negotiate alternative arrangements, rely more heavily on domestic chip designs from vendors such as Huawei and Biren, or accept higher latency by routing requests through data centres in China. None of those options matches the performance and cost profile of a regional deployment using Nvidia H100 or H200 GPUs, which remain the de facto standard for training and serving models at Hy4's scale.

The Open-Weight Economics

Chinese labs have converged on open-weight releases as a distribution strategy. By publishing model parameters under permissive licences, they lower the barrier for enterprises to experiment, fine-tune on proprietary data, and integrate inference into existing applications. The trade-off is that open weights do not generate direct API revenue; instead, the lab monetises through cloud compute, enterprise support contracts, and differentiated services such as safety filters or domain-specific fine-tunes.

Tencent's pricing for Hy4 sits in the middle of the market. DeepSeek V4 Pro offers off-peak rates that undercut most competitors, yet Tencent's position between off-peak and peak suggests confidence that Hy4's performance justifies a modest premium. The forthcoming Hy4 Flash, optimised for speed and cost efficiency, will likely target use cases where response time matters more than maximum reasoning depth, such as customer service chatbots, real-time translation, and content recommendation.

Alibaba's Qwen 3.8 Max, released in August, claimed to rank "second only to Fable 5," Anthropic's then-current flagship. Moonshot AI's Kimi K3.1 Pro and its Highspeed variant have posted strong results on agentic coding and frontend benchmarks, areas where developers can measure improvements directly through pass rates on programming challenges. Tencent enters a market in which open-weight models from domestic rivals already cover a wide performance band, and differentiation depends on niche strengths, pricing, and ecosystem integration.

What the Benchmarks Reveal

Hy4's strong showing in agentic coding and search tasks indicates that Tencent has invested in post-training techniques that improve tool-use reliability. Reinforcement learning from human feedback, supervised fine-tuning on curated agent traces, and synthetic data generation all contribute to a model's ability to handle multi-step workflows without hallucinating intermediate outputs or dropping context. Those capabilities are critical for enterprise adoption, where incorrect API calls or malformed queries can cascade into costly errors.

The weaker performance on AI-assisted research tasks suggests that Hy4's reasoning depth and factual grounding lag behind models that have seen larger or higher-quality pre-training corpora. Research workflows demand sustained coherence over long documents, precise citation of sources, and the ability to synthesise information across disparate domains. Closing that gap will require either more compute during pre-training, better data curation, or advances in post-training that teach the model to verify its own outputs before committing to an answer.

Tencent's candour about the gaps is notable. Many labs emphasise benchmark victories while downplaying weaknesses; acknowledging that "a considerable distance" remains between Hy4 and leading closed-source models sets realistic expectations for developers evaluating the system. It also signals that Tencent views the Hy series as a multi-year programme rather than a one-time release, with each iteration building on lessons learned from deployment and user feedback.