Tencent Open-Sources Hy4 Preview: 770B-Parameter Model Outperforms DeepSeek V4 Pro on Coding Benchmarks, but Over-Verification Slows Real-World Delivery

On August 28, 2026, Tencent's Hunyuan team officially released and open-sourced its next-generation flagship large language model, Hy4 Preview. The model adopts a MoE architecture with 770 billion total parameters (770B), activating 49B parameters per token, and a context window exceeding 1 million tokens. In the coding and software engineering domain, Hy4 Preview scored 85.4 on the Terminal Bench 2.1 command-line programming benchmark, surpassing DeepSeek V4 Pro and tying with Claude Opus 5; on the DeepSWE benchmark focused on real-world software engineering tasks, the score jumped sharply from the previous generation Hy3's 28.0 to 64.3. This marks the third node in the cadence of major version iterations roughly every two months since Tencent Hunyuan rebuilt its infrastructure in February 2026.

Architectural Trade-offs Behind the Parameter Doubling

The 770B scale represents a 2.6x increase in total parameters from the previous generation Hy3's 295B, with activated parameters also doubling from 28B to 49B. But what truly reflects the engineering trade-off is not the absolute scale but the activation ratio: out of the 770B total, only 49B is activated per inference, an activation rate of roughly 6.4%. This is the core logic of MoE sparse activation—using a large number of "dormant" experts in exchange for reasonable per-inference compute consumption.

The model itself was also used to optimize Tencent's own training and inference systems. According to Tencent's official disclosure, Hy4 Preview participated in the automated optimization of training methods, data strategies, and evaluation frameworks, improving end-to-end inference throughput by 31.8% over the baseline. This closed loop of "the model optimizing systems, the systems training the model" is one of the most technically sophisticated parts of this release, but the 31.8% figure comes from Tencent's internal measurements and has yet to be reproduced by third parties.

The training data construction strategy was collaboratively created by Tencent's internal experts across software engineering, gaming, finance, security, and other domains, and is continuously calibrated through deep co-design with products such as WorkBuddy. This strategy of "products feeding back into the model" is how Tencent monetizes its asset of massive real-world business scenarios.

Full-Spectrum Distribution Across Twelve Benchmarks

Looking only at the 85.4 on Terminal Bench 2.1 and the 64.3 on DeepSWE could easily lead to selective optimism. Examining the 12 benchmarks Tencent published as a whole reveals a more complete picture. According to explainx.ai's compilation, Hy4 Preview scored 65.7 on SWE-Bench Pro, 83.7 on the MCP-Atlas tool-calling evaluation, 92.3 on GPQA Diamond academic reasoning, and 74.1 on Toolathlon-Verified tool-calling marathon (according to Tencent Technology's report, surpassing Qwen 3.8 Max and GPT-5.6 Sol). On APEX-Agents (pass@1), which tests agents' one-shot success rate, Hy4 Preview scored 37.1, second only to Kimi K3's 37.2.

The highlight of this report card is its "lack of obvious weaknesses"—Tencent officially stated that the model did not rank last in any of the 12 benchmarks, even though its parameter scale is far smaller than some leading competitors. But this "balanced" characteristic is itself a signal: Hy4 Preview is closer to an "all-around practical performer" than a breakthrough outlier in any single direction.

Tencent also conducted an internal blind evaluation: 163 internal experts rated model outputs on 203 WorkBuddy engineering tasks using a 4-point scale. The results showed Hy4 Preview averaging 2.99, slightly ahead of Kimi K3's 2.94 and GLM-5.3's 2.92. According to explainx.ai's analysis, the 0.05-point lead is statistically close to noise, and the performance gap among the three models on similar productivity tasks has approached an "interchangeable" level.

Known Limitations: No Vision, Slow Delivery

The Tencent Hunyuan team proactively disclosed two known limitations at launch. First, Hy4 Preview currently does not support visual input—multimodal capabilities are absent and will need to be filled in by a future official release. Second, which is more noteworthy for actual users: the model exhibits an "over-verification tendency" in complex tasks—repeatedly verifying during task execution rather than delivering directly, resulting in notably longer overall response times. According to Tencent Technology's hands-on description, "the result is about to appear, but the model keeps verifying, causing users to wait longer." According to explainx.ai's records of OpenRouter real-world measurement data, under P50 latency the inference speed is approximately 36 tokens per second, with 3.19 seconds elapsing before output begins.

These two issues differ in nature. The vision gap is a functional shortfall that the company has explicitly committed to completing in the official release. But over-verification is an inference behavior issue, which Tencent has characterized as a "known issue of the preview version," attributed to room for improvement in the post-training stage. For enterprise scenarios requiring high-concurrency calls or sensitive to response latency, this is a risk point that needs to be quantified through actual stress testing.

Dual Signals in Pricing and Openness Strategy

Hy4 Preview's pricing continues Hunyuan's "inclusive accessibility" approach. According to prices announced by the Tencent Hunyuan team, API call fees are 6 RMB per million input tokens, 18 RMB per million output tokens, and 0.3 RMB per million tokens for cache hits. Converted to US dollars, the OpenRouter listed price is $0.834 per million input tokens and $2.501 per million output tokens.

The open-source release covers three platforms—HuggingFace, ModelScope, and GitCode—with fully public weights. API access channels include Tencent Cloud TokenHub and OpenRouter, and WorkBuddy and CodeBuddy offer a two-week free trial period. Meanwhile, the free access period for the previous generation Hy3 on the above platforms has been extended to September 30—an arrangement that gives developers a time window to test the old and new versions in parallel, and can also be understood as Tencent's proactive acknowledgment of Hy4 Preview's "preview status": the current version is not the final form, so there is no need to rush the switch.

Tencent's R&D spending this quarter grew 35% year-over-year to 27.3 billion RMB (approximately $4.1 billion), while capital expenditure grew 176% year-over-year to 52.8 billion RMB (approximately $7.9 billion). These two figures come from Tencent's quarterly report data as cited by finance.biggo.com, providing financial context for Hy4 Preview's compute investment—especially the near-doubling of capital expenditure, which points to large-scale GPU infrastructure expansion.

Competitive Landscape: A Three-Way Structure Takes Initial Shape

In terms of coding and agent capabilities, the arrival of Hy4 Preview has shifted the competitive landscape of domestic open-source models from a "DeepSeek solo lead" to a stage where multiple players are closing the gap. The 85.4 score on Terminal Bench 2.1 ties with Claude Opus 5 and surpasses DeepSeek V4 Pro. The significance of this number lies not in its precise ranking, but in the fact that it marks the first time a domestic model has entered the international first tier on this core coding benchmark.

For enterprise users, the choice logic offered by the current landscape can be summarized as follows: DeepSeek V4 Pro remains competitive in inference speed and certain general tasks; Kimi K3 is very close to Hy4 Preview in tool calling (Toolathlon) and agent one-shot success rate; GLM-5.3 trails Hy4 Preview by only 0.07 points in the internal blind evaluation. The three models are approaching a "performance ceiling" on text-only productivity tasks, and the differentiators are shifting to three dimensions: inference speed, deployment cost, and ecosystem integration.

For developers, here is the selection advice: if your current project relies on code generation and command-line tasks, Hy4 Preview's Terminal Bench and DeepSWE results warrant focused testing, but you must evaluate the latency impact of its over-verification issue under real workloads. If the project involves multimodal input, the current version is not usable—you will need to wait for the official release or choose an alternative. If cost-sensitive, the 6 RMB per million input token price is competitive for the parameter scale, but the true cost should be calculated together with the actual token consumption required to complete tasks (which is affected by over-verification).