Tencent’s AI large model iteration speed is astonishingly fast. The official version of Hunyuan Hy3 was only released in early July, and on the morning of August 28, the Tencent Hunyuan team again on XOfficial announcement Hy4 preview officially releases weights. 770B total parameters, 49B active parameters, 1M token context, Apache 2.0 license, on the same day. Hugging Face、GitHub andTencent Cloud 。
Tencent China launches the open-weight Hunyuan Hy4 preview model.
Parameter scale has doubled, and the architecture continues along the MoE route.
Hy4 preview is the third major version since Tencent Hunyuan rebuilt its training infrastructure in February this year. The previous generation, Hy3, had 295B total parameters with 21B activated; Hy4 preview has doubled both figures to 770B and 49B, and the context window has been expanded from 262K to 1M. The architecture remains a Mixture-of-Experts model: the backbone has 78 layers in total, with the first layer using a standard dense FFN and the remaining 77 layers adopting MoE routing. The Tencent Hunyuan team stated that the model has been comprehensively scaled up in terms of model size, context length, and data scale, with simultaneous improvements in both pretraining and post-training, solidifying its position in the top tier of open-source models.
The 1M context isn’t achieved by forcibly stretching it with rope-scaling. According to the model card, Hy4 preview’s long context is supported by sparse attention paired with an IndexCache design. The actual limit listed on OpenRouter is 1,048,576 tokens, with a maximum output of 64,000 tokens. For workflows that need to process an entire codebase or long documents, this window is significantly wider than most competitors.
12 benchmark tests comprehensively surpass Hy3.
Among the 12 benchmarks published by Tencent, Hy4 preview did not score the lowest on any of them. A few key numbers:
Terminal-Bench 2.1 (terminal command-line coding) scored 85.4, beating Hy3’s 70.8 by 14.6 points, surpassing DeepSeek V4 Pro and tying with Claude Opus 5 in the top tier. DeepSWE (real-world software engineering long tasks) jumped from Hy3’s 28.0 to 64.3, the most dramatic improvement across all four metrics. Toolathlon-Verified (tool-calling marathon) scored 74.1, ahead of Qwen 3.8 Max and GPT-5.6 Sol, and close to Kimi K3 and Claude Opus 5. APEX-Agents (pass@1, agent first-attempt success rate) scored 37.1, nearly matching Kimi K3’s 37.2.
The internal blind test provided another set of references. Tencent invited 163 internal experts to blindly evaluate 203 engineering tasks in the WorkBuddy environment. Hy4 preview scored an average of 2.99/4.00, narrowly edging out GLM 5.3 (2.92) and Kimi K3 (2.94). This data comes from in-house testing, so its reference value should be discounted, but at least it shows that the model’s performance in real engineering workflows is consistent with the direction of its benchmark scores.
From a parameter efficiency perspective, Hy4 preview’s 770B total parameters and 49B active parameters are far lower than competing models at the same level. OrcaRouter’s comparison article specifically notes that Hy4 preview’s score on Terminal-Bench has already surpassed DeepSeek V4 Pro, yet the latter has a significantly larger active parameter count. MoE routing allows the model to activate only a subset of experts during inference, so the actual computation is far below what the total parameter count suggests. Teams doing self-hosting can run a model with near-top-tier capability using fewer GPU resources.
Four Key Scenarios: Built for Productivity
Tencent defines Hy4 preview as “built for productivity,” with a focus on four scenarios. In software engineering, the model shows significant improvements in understanding, planning, debugging, and verification of long tasks, and also optimizes the visual aesthetics and interaction quality of front-end development. In office analysis, it enhances understanding of complex office environments and financial analysis capabilities, with a focus on optimizing data analysis and cross-file collaboration, able to go from information processing all the way to delivery of documents, spreadsheets, and presentations.
Game development is a relatively rare scenario that gets its own dedicated section. Hy4 preview can generate a playable prototype directly from a single requirements prompt, and iterate on complex game projects through multi-round interactions. For indie developers or small studios, this capability dramatically lowers the barrier to prototype validation: write a description of the requirements, the model directly produces a runnable version, then fine-tune parameters and mechanics based on actual feel.
In scientific research, the model has made progress in AI development, molecular dynamics simulations, condensed matter physics, and fundamental mathematics. The standout result is the three-dimensional Blaschke-Lebesgue geometry problem: Hy4 preview combined with Hyra pushed the volume lower bound from 0.380799 to 0.41104, leaving only a 2% gap from the 0.41986 given by the Meissner tetrahedron conjecture. While this result is not yet a final proof, it has advanced a geometry problem that had remained open for years to a point close to convergence.
Pricing and Open Strategy
API pricing continues Hunyuan’s consistent “high cost-performance” approach: 6 RMB per million input tokens (about NT$28), 18 RMB for output (about NT$84), and 0.3 RMB for cache hits (about NT$1.4). USD pricing on OpenRouter is $0.834/M for input, $2.501/M for output, and $0.042/M for cache reads.[6]Compared with open-source models of the same level, this price falls into the lowest tier for the 770B scale.
The licensing continues to use Apache 2.0, consistent with the open strategy of the official Hy3 release. When Hy3 transitioned from a restrictive community license to Apache 2.0 last month, the LocalLLaMA community’s reaction came down to one thing: it could finally be legally used in commercial and region-restricted scenarios. Hy4 preview launched directly under Apache 2.0, skipping that transition altogether. The model is simultaneously integrated into Tencent products including WorkBuddy, CodeBuddy (domestic and international versions), Yuanbao, and ima. Developers can experience it directly through these products or access it via API.
Known limitations and next steps
The Tencent Hunyuan team has themselves listed two known issues: a tendency toward overthinking and excessive self-verification on complex tasks, and currently no multimodal (vision) capabilities. The preview version is designed specifically for gathering feedback; the team hopes that real-world usage by developers will help bring substantial improvements to the official Hy4.
The model’s “self-evolution” is another detail worth noting in this release. Tencent disclosed a case of Hy4 preview participating in its own R&D pipeline: for the first time, the model played an active role in the automated optimization of training methods, data strategies, evaluation frameworks, and underlying operators—proposing solutions, executing experiments, and continuing to iterate based on results, forming a preliminary recursive self-improvement loop.[3]This mechanism is only in its early stages on Hy4 preview, but the direction points toward a future where models no longer rely entirely on human engineers manually tuning parameters.
Infrastructure investment is the supporting foundation behind this. Tencent’s R&D spending in this quarter grew 35% year-over-year to RMB 27.3 billion (approximately NT$128 billion), while capital expenditure surged 176% to RMB 52.8 billion (approximately NT$246 billion). Management’s roadmap laid out during the earnings call is “build proprietary models first, then develop applications, and finally lease computing power externally,” with the goal of building a scaled native AI business.
From the infrastructure rebuild in February to now, Hunyuan has iterated a major version every two months on average. At this pace, the official Hy4 release isn’t far off. The Hy4 preview’s 770B MoE architecture, 1M context window, and Apache 2.0 license put it at the forefront of open-source models in both parameter scale and measured capability. What remains to be seen is feedback from the developer community in real-world scenarios, and how the official version rounds out multimodal capabilities and efficiency.