Tencent Releases Open-Source Hy4 AI Model for Coding, Research and Office Work
01 Event
Tencent has released and open-sourced Hy4 preview, a new large language model designed for practical work including coding, office tasks and scientific research. Tencent says the model has 770 billion total parameters, 49 billion active parameters and a context window exceeding one million tokens.
The model is available through Tencent products including WorkBuddy and CodeBuddy and can also be accessed through APIs and third-party model services. Tencent is offering temporary free access through some of its applications.
02 What Changed?
Chinese AI companies have increasingly moved from simply matching U.S. chatbots toward competing on open models, cost and developer access. Hy4 continues that trend by giving developers another large model they can inspect, deploy and integrate rather than relying exclusively on a closed hosted service.
The long context window is especially relevant for coding repositories, research papers and large business document collections where a model may need to reason across far more information than fits in a conventional prompt.
03 Why It Matters
Open-source competition can put downward pressure on AI pricing. Businesses that can host or access capable open models have more negotiating leverage when comparing proprietary services. It also reduces dependence on a small number of U.S. providers.
For developers, model diversity matters because different systems perform better on different workloads. A model optimized for coding or long documents may provide better economics than a general-purpose premium chatbot.
04 What It Means for You
Businesses should evaluate Hy4 on their own workloads rather than relying solely on benchmark claims. Test accuracy, latency, cost, language performance, security requirements and integration effort.
Developers should also distinguish between “open source” and “free to operate.” Large models still require significant computing resources. A downloadable model can avoid per-token API pricing while creating infrastructure and engineering costs of its own.
For ordinary users, stronger open models can eventually mean more competition among AI assistants and lower prices for AI-powered software.
05 Numbers + Context
Tencent lists 770 billion total parameters and 49 billion active parameters for Hy4 preview, indicating a mixture-of-experts architecture in which only part of the model is active for a given request. The context window exceeds one million tokens, enabling unusually large inputs.
Tencent says the model is available globally through WorkBuddy and CodeBuddy and through API channels including Tencent Cloud TokenHub and OpenRouter.
Related Earnyx coverage: Read why companies struggle to move AI beyond pilot projects and how AI models and hardware are expanding into industrial systems.
06 Earnyx Takeaway
The important story is not another benchmark leaderboard. It is the continued commoditization of capable AI. As more companies release competitive open models, businesses gain alternatives to expensive closed systems.
That does not mean the cheapest model is automatically the best. Reliability, support, data governance and deployment cost can outweigh token price. The right comparison is total cost per successful task.
For AI buyers, competition is valuable. Avoid designing workflows so tightly around one provider that switching becomes prohibitively expensive. Portable prompts, standard APIs and independent data storage preserve leverage as the model market changes.
Benchmark results should be treated as a starting point rather than a purchasing decision. Model vendors choose tests that highlight different strengths, and small score differences may have little relationship to a company’s actual workload. A finance team extracting data from spreadsheets, for example, cares more about error rates and auditability than performance on a coding benchmark.
The one-million-token context window also needs practical testing. Being able to accept a very large input does not guarantee the model will use every part of that input equally well. Long-context evaluations should check whether important details buried deep in documents are retrieved accurately and whether costs and latency remain acceptable.
Open models can provide stronger data control when deployed inside a company’s own environment. That can be attractive for sensitive research or proprietary code. But self-hosting transfers responsibility for security, updates, monitoring and infrastructure to the customer. The absence of an API bill does not make the deployment free.
Competition from Tencent and other Chinese developers also has geopolitical implications. Governments are increasingly concerned about where AI models are developed, where data is processed and what export controls apply to advanced chips. Global businesses may eventually maintain different AI stacks for different regions.
Developers should therefore evaluate licensing terms carefully. “Open” can describe several different arrangements, from downloadable weights with restrictions to permissive open-source licenses. The exact license determines whether a model can be modified, redistributed or used commercially.
For smaller businesses, the best outcome from the open-model race may simply be lower prices. Even companies that never deploy Hy4 can benefit if its availability forces competing providers to improve models, expand free tiers or reduce API charges.
The market is moving quickly enough that long-term lock-in deserves a real cost in purchasing decisions. A slightly cheaper tool today may become expensive if all prompts, data pipelines and workflows must be rebuilt to switch later. Standard interfaces and portable evaluation sets preserve flexibility.
Teams should maintain a repeatable evaluation set made from real tasks before switching models. Running the same prompts, documents and code problems across several providers makes quality differences measurable instead of subjective. Cost should then be compared per successful result, because a cheaper model that requires repeated corrections can cost more in staff time.
Hy4 also shows why AI procurement should be revisited regularly. Model capability and pricing can change materially within months. Contracts and architecture that allow periodic competition can capture those improvements instead of locking a business into assumptions made when the market looked very different.
The winner for a particular company may therefore change repeatedly as models improve.
Regular testing keeps that decision grounded in current performance rather than brand reputation.
Source: Tencent’s August 28, 2026 announcement of Tencent Hy4 preview and related reporting.