Tencent’s Hy4 Preview is Tencent’s largest open-weight language-model release to date and its clearest attempt to connect a frontier-scale model with workplace and coding products. Released on August 28, 2026, it targets software engineering, research, financial analysis, and other agentic productivity tasks. 112

The headline specifications are substantial: 770 billion total parameters, 49 billion active parameters per token, and a context window exceeding 1 million tokens. But the more important conclusion is qualified: Hy4 looks like a top-tier Chinese open model for coding and productivity, while its claims of superiority still rely heavily on Tencent’s own evaluations and await broader independent testing. 18

What is Tencent Hy4 Preview?

Hy4 Preview is a Mixture-of-Experts (MoE) large language model. Instead of using every parameter for every token, its routing system activates a smaller subset of the network for each request. That lets Tencent build a much larger total model without imposing the full theoretical parameter cost on every token.

Tencent describes Hy4 as a model for agentic work rather than a general-purpose chatbot alone. Its intended uses include coding, office productivity, research, and financial analysis, with integrations designed to put the model directly inside software-development and workplace workflows. 18

“Open source” is best understood here as open weights and permissive deployment support. Tencent released the model under the Apache 2.0 license, alongside serving, fine-tuning, and quantization guidance. 812

Hy4 Preview specifications and architecture

The main specifications are:

• Total parameters: 770 billion
• Active parameters per token: 49 billion
• Context window: More than 1 million tokens
• Architecture: Mixture of Experts
• License: Apache 2.0
• Model formats: BF16 and FP8 versions
• Serving support: vLLM and SGLang recipes, with an OpenAI-compatible API path 812

The model’s backbone contains 78 layers. Tencent says the first uses a dense feed-forward network, while the remaining 77 use MoE routing with 256 routed experts and a shared expert. Each token selects eight routed experts alongside the shared expert. Hy4 also includes a native multi-token-prediction layer intended to support speculative decoding. 8

Its attention design uses Gated DeepSeek Sparse Attention with IndexCache for reusing sparse attention indices across layers. Tencent says the architecture was inspired in part by approaches associated with DeepSeek and GLM. 8

These design choices matter because Hy4’s appeal is not simply its parameter count. The model combines a very large capacity ceiling with sparse routing, long-context processing, and mechanisms intended to reduce inference overhead.

Hy4 versus Hy3: what changed?

Hy4 is a substantial expansion of Tencent’s previous Hy3 model:

Specification	Hy3	Hy4 Preview	Change
Total parameters	295B	770B	About 2.6× larger
Active parameters per token	21B	49B	About 2.3× more active compute
Context window	262K tokens	More than 1M tokens	Roughly 4× larger
Model design	MoE	Expanded MoE with sparse-attention changes	Higher capacity and longer context

Hy4 therefore increases both the total model capacity and the amount of active computation used for each token. Tencent also says it expanded training data and made substantial changes to pre-training and post-training. 814

The context-window increase may be especially important for practical users. A window exceeding 1 million tokens can help with large codebases, lengthy research collections, and complex workplace tasks—although actual usefulness will depend on retrieval quality, latency, and how well the model maintains accuracy across long inputs.

How does Hy4 compare with Chinese competitors?

Tencent’s clearest published comparison is an internal blind side-by-side evaluation. In that test, 163 internal experts assessed outputs on 203 engineering tasks. Hy4 Preview received an average score of 2.99 out of 4, compared with 2.92 for Zhipu’s GLM-5.3 and 2.94 for Moonshot’s Kimi K3. 8

That result places Hy4 slightly ahead in Tencent’s test, but it should not be treated as a neutral, comprehensive ranking. The evaluators were Tencent’s internal experts, and the test focused on engineering tasks. Reporting on other benchmarks shows Hy4’s results varied against domestic competitors such as Alibaba’s Qwen models and DeepSeek. 113

Some secondary reports also cite large gains over Hy3 on coding benchmarks, including Terminal-Bench 2.1 and DeepSWE. Those figures are useful indicators of Tencent’s claimed progress, but they should be interpreted cautiously until benchmark conditions and results are independently reproduced. 1718

The defensible comparison is that Hy4 belongs in the top tier of Chinese open-weight models for coding and productivity workloads. The available evidence does not establish that it decisively outperforms every competing model across every task.

Where can you use Hy4 Preview?

Tencent has made Hy4 available through both downloadable weights and hosted products.

Downloadable weights and deployment

The model and its FP8 version are available through Hugging Face, ModelScope, GitCode, and CNB. Tencent provides deployment recipes for vLLM and SGLang, an OpenAI-compatible local API route, fine-tuning instructions, and quantization tools. 812

This makes Hy4 more accessible to developers who want to experiment with the weights or build a customized serving stack. However, “open weight” does not mean lightweight. Tencent’s reference FP8 configuration uses eight-way tensor parallelism, and serving a 770B model remains a serious multi-GPU infrastructure project. 8

Tencent applications and hosted APIs

Users can access Hy4 through WorkBuddy, CodeBuddy, Yuanbao, and ima. Tencent also lists access through Tencent Cloud TokenHub and OpenRouter. At launch, Tencent offered two weeks of free access through WorkBuddy and CodeBuddy, while access to the previous Hy3 model was extended through September 30. 12

A third-party report listed OpenRouter input pricing at approximately $0.834 per million tokens. That figure should be treated as channel-specific rather than a permanent Tencent list price, since pricing and availability can vary by provider and geography. 3

Why Tencent’s product distribution matters

Hy4 is closely tied to Tencent’s application strategy. Tencent says the model was co-designed with WorkBuddy and CodeBuddy, using task data and feedback from internal engineers, game developers, finance analysts, and security specialists. 8

That approach gives Tencent a potential advantage beyond model architecture: it can test the model inside products used for real coding and office workflows, then feed product feedback into later iterations. The same distribution network also gives users multiple ways to encounter the model without downloading and operating a 770B system themselves.

There are early signs that Tencent’s broader Hunyuan distribution can generate significant usage. Tencent reported that putting Hunyuan 3 into production increased average daily token use by roughly six times compared with its preview phase. This is evidence of rapid usage growth, not independent proof of user satisfaction or long-term retention.

Tencent is also spending heavily to support this strategy. Facilities-and-equipment capital expenditure reached RMB 52.784 billion in the second quarter of 2026, approximately 2.8 times the year-earlier level, with management prioritizing AI infrastructure and investment related to WorkBuddy.

Tencent has described WorkBuddy as China’s most widely used productivity AI-agent service, but that is a company claim rather than an independently audited market-share measurement. 4

Current limitations

Hy4 Preview is not presented as a finished model. Tencent acknowledges that it can spend longer than necessary reasoning through difficult tasks and may over-verify its own work. The result can be higher latency and inference cost. 8

The model’s scale creates a second practical limitation. Although MoE routing means only 49B parameters are active for an individual token, the full model still has to be stored and served. Organizations considering self-hosting need substantial memory, networking, and multi-GPU capacity.

The benchmark evidence also has limits. Tencent’s internal evaluation shows a narrow lead over GLM-5.3 and Kimi K3 in engineering tasks, while broader comparative results vary by benchmark. Early rankings should therefore be treated as directional rather than definitive. 8

Finally, long context is not automatically equivalent to reliable long-context reasoning. Hy4’s 1-million-token window is a major specification advantage, but real-world performance will depend on the task, prompt structure, retrieval strategy, and serving configuration.

What improvements are likely next?

Tencent’s stated direction is clearer than a formal release roadmap. The company points to additional pre-training and post-training, faster and less repetitive reasoning, continued feedback from WorkBuddy and CodeBuddy, and compression or quantization work to lower deployment costs. 8

Those improvements would address Hy4’s central trade-off: it offers more capacity and longer context than Hy3, but the preview label signals that speed, efficiency, and production reliability are still being refined.

Bottom line

Hy4 Preview is a major step up from Hy3: it expands from 295B to 770B total parameters, raises active parameters from 21B to 49B, and grows the context window from 262K to more than 1 million tokens. It is also unusually well connected to Tencent’s own coding, office, cloud, and consumer-product ecosystem. 812

For developers, the Apache 2.0 weights and deployment tooling make Hy4 worth evaluating. For organizations, hosted access through Tencent’s products and APIs may be more practical than self-hosting. But the right current verdict is measured rather than absolute: Hy4 looks like a powerful top-tier Chinese open-weight model for coding and agentic productivity, while independent benchmarks and sustained production use will determine whether its early promise translates into a durable lead.

Tencent’s Hy4 Preview is Tencent’s largest open-weight language-model release to date and its clearest attempt to connect a frontier-scale model with workplace and coding products. Released on August 28, 2026, it targets software engineering, research, financial analysis, and other agentic productivity tasks. 112

The headline specifications are substantial: 770 billion total parameters, 49 billion active parameters per token, and a context window exceeding 1 million tokens. But the more important conclusion is qualified: Hy4 looks like a top-tier Chinese open model for coding and productivity, while its claims of superiority still rely heavily on Tencent’s own evaluations and await broader independent testing. 18

What is Tencent Hy4 Preview?

Hy4 Preview is a Mixture-of-Experts (MoE) large language model. Instead of using every parameter for every token, its routing system activates a smaller subset of the network for each request. That lets Tencent build a much larger total model without imposing the full theoretical parameter cost on every token.

Tencent describes Hy4 as a model for agentic work rather than a general-purpose chatbot alone. Its intended uses include coding, office productivity, research, and financial analysis, with integrations designed to put the model directly inside software-development and workplace workflows. 18

“Open source” is best understood here as open weights and permissive deployment support. Tencent released the model under the Apache 2.0 license, alongside serving, fine-tuning, and quantization guidance. 812

Hy4 Preview specifications and architecture

The main specifications are:

• Total parameters: 770 billion
• Active parameters per token: 49 billion
• Context window: More than 1 million tokens
• Architecture: Mixture of Experts
• License: Apache 2.0
• Model formats: BF16 and FP8 versions
• Serving support: vLLM and SGLang recipes, with an OpenAI-compatible API path 812

The model’s backbone contains 78 layers. Tencent says the first uses a dense feed-forward network, while the remaining 77 use MoE routing with 256 routed experts and a shared expert. Each token selects eight routed experts alongside the shared expert. Hy4 also includes a native multi-token-prediction layer intended to support speculative decoding. 8

Its attention design uses Gated DeepSeek Sparse Attention with IndexCache for reusing sparse attention indices across layers. Tencent says the architecture was inspired in part by approaches associated with DeepSeek and GLM. 8

These design choices matter because Hy4’s appeal is not simply its parameter count. The model combines a very large capacity ceiling with sparse routing, long-context processing, and mechanisms intended to reduce inference overhead.

Hy4 versus Hy3: what changed?

Hy4 is a substantial expansion of Tencent’s previous Hy3 model:

Specification	Hy3	Hy4 Preview	Change
Total parameters	295B	770B	About 2.6× larger
Active parameters per token	21B	49B	About 2.3× more active compute
Context window	262K tokens	More than 1M tokens	Roughly 4× larger
Model design	MoE	Expanded MoE with sparse-attention changes	Higher capacity and longer context

Hy4 therefore increases both the total model capacity and the amount of active computation used for each token. Tencent also says it expanded training data and made substantial changes to pre-training and post-training. 814

The context-window increase may be especially important for practical users. A window exceeding 1 million tokens can help with large codebases, lengthy research collections, and complex workplace tasks—although actual usefulness will depend on retrieval quality, latency, and how well the model maintains accuracy across long inputs.

How does Hy4 compare with Chinese competitors?

Tencent’s clearest published comparison is an internal blind side-by-side evaluation. In that test, 163 internal experts assessed outputs on 203 engineering tasks. Hy4 Preview received an average score of 2.99 out of 4, compared with 2.92 for Zhipu’s GLM-5.3 and 2.94 for Moonshot’s Kimi K3. 8

That result places Hy4 slightly ahead in Tencent’s test, but it should not be treated as a neutral, comprehensive ranking. The evaluators were Tencent’s internal experts, and the test focused on engineering tasks. Reporting on other benchmarks shows Hy4’s results varied against domestic competitors such as Alibaba’s Qwen models and DeepSeek. 113

Some secondary reports also cite large gains over Hy3 on coding benchmarks, including Terminal-Bench 2.1 and DeepSWE. Those figures are useful indicators of Tencent’s claimed progress, but they should be interpreted cautiously until benchmark conditions and results are independently reproduced. 1718

The defensible comparison is that Hy4 belongs in the top tier of Chinese open-weight models for coding and productivity workloads. The available evidence does not establish that it decisively outperforms every competing model across every task.

Where can you use Hy4 Preview?

Tencent has made Hy4 available through both downloadable weights and hosted products.

Downloadable weights and deployment

The model and its FP8 version are available through Hugging Face, ModelScope, GitCode, and CNB. Tencent provides deployment recipes for vLLM and SGLang, an OpenAI-compatible local API route, fine-tuning instructions, and quantization tools. 812

This makes Hy4 more accessible to developers who want to experiment with the weights or build a customized serving stack. However, “open weight” does not mean lightweight. Tencent’s reference FP8 configuration uses eight-way tensor parallelism, and serving a 770B model remains a serious multi-GPU infrastructure project. 8

Tencent applications and hosted APIs

Users can access Hy4 through WorkBuddy, CodeBuddy, Yuanbao, and ima. Tencent also lists access through Tencent Cloud TokenHub and OpenRouter. At launch, Tencent offered two weeks of free access through WorkBuddy and CodeBuddy, while access to the previous Hy3 model was extended through September 30. 12

A third-party report listed OpenRouter input pricing at approximately $0.834 per million tokens. That figure should be treated as channel-specific rather than a permanent Tencent list price, since pricing and availability can vary by provider and geography. 3

Why Tencent’s product distribution matters

Hy4 is closely tied to Tencent’s application strategy. Tencent says the model was co-designed with WorkBuddy and CodeBuddy, using task data and feedback from internal engineers, game developers, finance analysts, and security specialists. 8

That approach gives Tencent a potential advantage beyond model architecture: it can test the model inside products used for real coding and office workflows, then feed product feedback into later iterations. The same distribution network also gives users multiple ways to encounter the model without downloading and operating a 770B system themselves.

There are early signs that Tencent’s broader Hunyuan distribution can generate significant usage. Tencent reported that putting Hunyuan 3 into production increased average daily token use by roughly six times compared with its preview phase. This is evidence of rapid usage growth, not independent proof of user satisfaction or long-term retention.

Tencent is also spending heavily to support this strategy. Facilities-and-equipment capital expenditure reached RMB 52.784 billion in the second quarter of 2026, approximately 2.8 times the year-earlier level, with management prioritizing AI infrastructure and investment related to WorkBuddy.

Tencent has described WorkBuddy as China’s most widely used productivity AI-agent service, but that is a company claim rather than an independently audited market-share measurement. 4

Current limitations

Hy4 Preview is not presented as a finished model. Tencent acknowledges that it can spend longer than necessary reasoning through difficult tasks and may over-verify its own work. The result can be higher latency and inference cost. 8

The model’s scale creates a second practical limitation. Although MoE routing means only 49B parameters are active for an individual token, the full model still has to be stored and served. Organizations considering self-hosting need substantial memory, networking, and multi-GPU capacity.

The benchmark evidence also has limits. Tencent’s internal evaluation shows a narrow lead over GLM-5.3 and Kimi K3 in engineering tasks, while broader comparative results vary by benchmark. Early rankings should therefore be treated as directional rather than definitive. 8

Finally, long context is not automatically equivalent to reliable long-context reasoning. Hy4’s 1-million-token window is a major specification advantage, but real-world performance will depend on the task, prompt structure, retrieval strategy, and serving configuration.

What improvements are likely next?

Tencent’s stated direction is clearer than a formal release roadmap. The company points to additional pre-training and post-training, faster and less repetitive reasoning, continued feedback from WorkBuddy and CodeBuddy, and compression or quantization work to lower deployment costs. 8

Those improvements would address Hy4’s central trade-off: it offers more capacity and longer context than Hy3, but the preview label signals that speed, efficiency, and production reliability are still being refined.

Bottom line

Hy4 Preview is a major step up from Hy3: it expands from 295B to 770B total parameters, raises active parameters from 21B to 49B, and grows the context window from 262K to more than 1 million tokens. It is also unusually well connected to Tencent’s own coding, office, cloud, and consumer-product ecosystem. 812

For developers, the Apache 2.0 weights and deployment tooling make Hy4 worth evaluating. For organizations, hosted access through Tencent’s products and APIs may be more practical than self-hosting. But the right current verdict is measured rather than absolute: Hy4 looks like a powerful top-tier Chinese open-weight model for coding and agentic productivity, while independent benchmarks and sustained production use will determine whether its early promise translates into a durable lead.