Tencent has launched Hy4 Preview, a low-cost, open-weight artificial intelligence model designed to compete with leading systems from OpenAI, Anthropic, DeepSeek, Zhipu AI and Moonshot AI. 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, and it comes with API pricing that undercuts much of the competition. It sharpens a contest increasingly about openness and cost as much as raw capability.
Hy4 Preview Specifications
The technical case for Hy4 Preview rests on scale combined with efficiency. The model has 770 billion total parameters but activates only 49 billion parameters per token, using a Mixture-of-Experts design in which only part of the model works on each token. It supports a one-million-token context window, enough to hold very long documents, codebases or conversations in a single session. The architecture runs to 78 layers, including 77 Mixture-of-Experts layers with 256 routed experts and one shared expert per layer, activating eight routed experts per token. It also uses Gated DeepSeek Sparse Attention and identity Hyper-Connections to improve information processing, with multi-token prediction as an additional layer. Released on August 28, 2026, under the Apache 2.0 licence, it is, by Tencent’s own description, its largest generation-over-generation improvement so far.
How Hy4 Compares With Rival Models
Tencent evaluated Hy4 Preview against several leading models using software engineering, agent-based and reasoning benchmarks, including Claude Opus 5, GPT-5.6 Sol, DeepSeek V4 Pro, GLM 5.3 and Kimi K3. The comparisons are Tencent-reported results, not an independent ranking, and some rivals carry two reported scores under different evaluation configurations. Even with that caveat, the pattern is interesting. Hy4 Preview scored 82.9 on SWE-bench Multilingual, ahead of the listed scores for DeepSeek V4 Pro, GLM 5.3, Kimi K3 and GPT-5.6 Sol, although Claude Opus 5 performed better on that test. It scored 64.0 on SWE Atlas Codebase Q&A, outperforming all five competitors shown, and 57.8 on test writing and 53.3 on refactoring. On Terminal-Bench 2.1 it scored 85.4, although several competitors reported higher results. The picture is of a model competitive across demanding programming tasks rather than strongest in every category.
Hy4 Against Hy3
The comparison with Tencent’s own previous generation is where the progress looks most dramatic. On SWE-bench Multilingual, Hy4 Preview improved from 75.8 to 82.9, and on SWE-bench Pro from 57.9 to 65.7. One of the largest jumps appeared on DeepSWE, where the score rose from 28.0 to 64.3, while SWE Atlas Codebase Q&A improved from 30.8 to 64.0 and ProgramBench from 3.0 to 17.5. CyberGym rose from 51.8 to 78.4, Toolathlon-Verified from 56.2 to 74.1 and JobBench from 34.6 to 61.7. Tencent credits larger-scale training and greater use of real-world engineering tasks.
Performance in Coding, Office Work and Research
Hy4 Preview also performed competitively on benchmarks designed to measure searching for information, working with documents and using external tools. On OneMillionBench with tools it scored 65.4, up from 51.5 for Hy3, while WideSearch rose to 83.9 and OfficeQA Pro to 66.2. On MCP-Atlas it reached 83.7, and on WorkspaceBench 60.2, although GLM 5.3 and Kimi K3 remained ahead in some agent and workspace evaluations. Tencent says the model delivers significant improvements in software engineering, office productivity, game development and scientific research.
Pricing and Access
Where Hy4 Preview may make its biggest impact is price. API pricing starts at $0.042 per million cached input tokens, while regular input costs $0.834 and output costs $2.501 per million tokens. Those figures are far below the prices charged by many proprietary frontier models, and they matter because token costs determine whether a business can afford to build AI into products used by thousands or millions of people. Access is through Tencent Cloud and OpenRouter, as well as Tencent’s AI applications such as CodeBuddy, WorkBuddy and Yuanbao, while developers can download the weights through supported open-source repositories including Hugging Face, ModelScope, GitCode and CNB. Open weights and low API prices give developers two routes in: run it themselves, or call it cheaply through the cloud.
Limitations to Keep in Mind
Tencent itself acknowledges drawbacks. The company notes excessive reasoning time and repeated self-checking, problems that can make the model slower and more verbose than users expect. The benchmark results, too, come with a health warning: they are Tencent’s own figures, and independent evaluation may tell a different story on some tasks. Buyers and developers should therefore treat Hy4 Preview as a strong, low-cost contender whose real-world performance still needs to be proven outside Tencent’s test suite.
The wider significance of Hy4 Preview lies in what it says about the direction of the AI industry. Open-weight models from China, including DeepSeek, GLM and Kimi, have already forced Western companies to justify premium prices, and Tencent’s entry adds another powerful, cheaply priced option. If it performs well independently, it will pressure rivals on price and give smaller companies frontier-class capability without frontier-class bills. That is a meaningful shift, and it is why a model launch from Tencent deserves attention well beyond the AI research community.
