Tencent Unveils Its Most Powerful AI Model at Extremely Low Cost

Tencent has introduced Hy4 Preview, a low-cost, open-weight AI model designed to compete with leading systems from OpenAI, Anthropic, DeepSeek, Zhipu AI, and Moonshot AI.

Key points

• Terminal-Bench 2.1 85.4 87.9 / 80.3 88.2 / 88.3 88.3 / 85.7 88.8 / 88.3 86.7 / 85.4.
• DeepSWE 64.3 62.7 / 58.8 66.9 / 68.1 67.5 / 74.0 72.7 / 68.9 68.8 / 74.7.
• SWE-bench Pro 65.7 60.3 64.6 63.3 64.6 / 60.5 79.2 / 79.9.
• SWE-bench Multilingual 82.9 77.3 81.3 80.8 74.1 89.5 / 85.8.
• SWE Atlas: Test Writing 57.8 45.6 49.6 35.6 49.6 68.7.

Unlike proprietary models such as GPT-5.6 and Claude Opus 5, Hy4 Preview offers publicly available model weights that developers can download and deploy independently. While regular input costs $0.834 and output costs $2.501 per million tokens, it additionally comes with relatively affordable API pricing, starting at $0.042 per million cached input tokens.

Meanwhile, the model has 770 billion total parameters, activates 49 billion parameters per token, and supports a one-million-token context window. Tencent notes it delivers significant improvements in software engineering, office productivity, game development, and scientific research.

According to Tencent's published benchmarks, Hy4 Preview outperforms a number of competing frontier models in selected coding and productivity tests. Nevertheless, Claude Opus 5, GPT-5.6 Sol, and other rivals remain ahead in some areas.

Published on August 28, 2026, Hy4 Preview is available through Tencent Cloud, OpenRouter, and Tencent's AI applications. Developers can additionally download its model weights through backed open-source repositories.

Hy4 Preview uses a Mixture-of-Experts (MoE) architecture. It activates only part of the model's total parameters for each token.

For context, the model contains 78 layers, including 77 MoE layers with 256 routed experts and one shared expert per layer.

It additionally uses Gated DeepSeek Sparse Attention and identity Hyper-Connections to improve information processing.

Google Just Created AI Talk From Space for the First Time Ever. Developer Tencent Hy Team. Total parameters 770 billion. Active parameters 49 billion per token. Context window 1 million tokens. Routed experts per MoE layer 256. Shared experts per MoE layer 1. Activated routed experts 8 per token. Attention mechanism Gated DeepSeek Sparse Attention. Residual architecture Identity Hyper-Connections. Extra prediction layer Multi-token prediction. Model versions Standard and FP8. License Apache 2.0. Release date August 28, 2026.

Tencent notes it climbed model size, context length, and training data while expanding both pre-training and post-training.

Notably, the firm describes Hy4 Preview as its largest generation-over-generation improvement so far. Hy4 Preview vs Major Models.

Tencent evaluated Hy4 Preview against a number of leading AI models using software engineering, agent-based, and reasoning benchmarks.

For context, the comparisons include Claude Opus 5, GPT-5.6 Sol, DeepSeek V4 Pro, GLM 5.3, and Kimi K3. Higher scores indicate better performance on each benchmark.

Benchmark Hy4 Preview DeepSeek V4 Pro GLM 5.3 Kimi K3 GPT-5.6 Sol Claude Opus 5.

SWE Atlas: Codebase Q&A 64.0 53.4 55.8 35.2 58.1 57.5.

SWE Atlas: Refactoring 53.3 48.6 51.9 37.4 52.4 60.0.

Some rivals have two documented scores under different evaluation configurations. These are retained as published. The comparisons are Tencent-reported results, not an independent ranking.

Hy4 achieved 82.9 on SWE-bench Multilingual, exceeding the listed scores for DeepSeek V4 Pro, GLM 5.3, Kimi K3, and GPT-5.6 Sol. Nevertheless, Claude Opus 5 performed better on that test.

Hy4 additionally achieved a score of 64.0 on SWE Atlas Codebase Q&A, outperforming all five competitors shown in the table.

Although a number of competitors noted higher results, on Terminal-Bench 2.1, it scored 85.4.

These figures indicate that Hy4 is competitive throughout demanding programming tasks, rather than the strongest model in every category. Tencent Hy4 vs Hy3: Improvements. Tencent additionally compared its new model with the previous Hy3 generation.

Notably, the results show improvements in coding, tool employ, research, and long-context processing. Benchmark Hy3 Hy4 Preview. SWE-bench Multilingual 75.8 82.9. SWE-bench Pro 57.9 65.7. DeepSWE 28.0 64.3. SWE Atlas: Codebase Q&A 30.8 64.0. SWE Atlas: Test Writing 35.9 57.8. SWE Atlas: Refactoring 32.9 53.3. Terminal-Bench 2.1 70.8 85.4. ProgramBench 3.0 17.5. CyberGym 51.8 78.4. Toolathlon-Verified 56.2 74.1. APEX-Agents 24.4 37.1. JobBench 34.6 61.7.

One of the largest improvements appeared on DeepSWE, where the score rose from 28.0 to 64.3.

In practice, the model additionally improved from 30.8 to 64.0 on SWE Atlas Codebase Q&A and from 3.0 to 17.5 on ProgramBench.

Tencent notes these gains resulted from larger-scale training and greater apply of real-world engineering tasks.

For context, the model additionally performed competitively on benchmarks designed to measure its ability to search information, work with documents, and apply external tools. Benchmark Hy3 Hy4 Preview GLM 5.3 Kimi K3. OneMillionBench (with tools) 51.5 65.4 64.5 63.5. WideSearch 81.9 83.9 83.2 81.0. OfficeQA Pro 54.1 66.2 66.2 65.4. MCP-Atlas 75.0 83.7 81.9 84.2 / 82.8.

Toolathlon-Verified 56.2 74.1 73.0 / 73.8 76.5 / 74.7. APEX-Agents 24.4 37.1 38.1 41.0 / 37.2. JobBench 34.6 61.7 58.2 54.3 / 54.7. WorkspaceBench 58.2 60.2 68.2 65.0.

Meanwhile, the results show that Hy4 improved significantly over Hy3 in document-heavy tasks and multi-step workflows.

While its OneMillionBench result improved from 51.5 to 65.4, on JobBench, its score rose from 34.6 to 61.7.

Nevertheless, GLM 5.3 and Kimi K3 remained ahead in some agent and workspace evaluations. Tests Against GLM and Kimi.

Tencent conducted a separate blind evaluation involving 163 internal experts and 203 software engineering tasks.

Participants assessed outputs produced by Hy4 Preview, GLM 5.3, and Kimi K3.

Hy4 received an average score of 2.99 out of 4, compared with 2.92 for GLM 5.3 and 2.94 for Kimi K3. Metric Hy4 vs GLM 5.3 Hy4 vs Kimi K3. Hy4 average score 2.99 2.99. Rival average score 2.92 2.94. Hy4 win rate 46.8% 51.2%. Tie rate 12.8% 7.9%. Hy4 loss rate 40.4% 40.9%.

Although the differences in average scores were small, hy4 performed slightly better in both comparisons.

Meanwhile, the tests were conducted internally by Tencent, so they should not be treated as independent confirmation of overall model superiority. Hy4 Is Designed for Coding, Office Work, Games and Research.

Taken together, the developments around tencent Unveils Its Most Powerful AI Model at Extremely Low Cost point to a situation that is still moving, and the coming days should bring more clarity.