腾讯混元发布并开源 Hy4 preview

2026 年 8 月 28 日，腾讯混元发布并开源 Hy4 preview。770B 总参、49B 激活的 MoE，上下文突破 1M，Apache 2.0 协议——这些数字把混元从国产第二梯队推进了第一梯队。但发布材料里最值得读的不是参数，而是一句话：「初步的递归自我改进闭环」。

先看规格。Hy4 preview 相比上代 Hy3（295B/21B、256K 上下文）规模大幅提升：总参 770B（Hugging Face 显示 780B，含内置 MTP 投机解码层）、激活 49B，上下文突破 1M。内部盲测（163 名专家、203 个工程任务）均分 2.99/4.00，略优于 GLM-5.3（2.92）与 Kimi K3（2.94）——需要带着厂商自述的边界读，但方向是清楚的：混元进入了国产开源旗舰的正面竞争。自述 DeepSWE 64.3（上代 Hy3 为 28.0）、Terminal Bench 2.1 达 85.4，定位「为生产力而生」，聚焦软件工程、办公分析、游戏开发与科学研究。

「递归自我改进」是这次发布最特别的叙事。官方称 Hy4 preview 首次参与自身研发全链路：训练方法、数据策略、评估体系、底层算子自动优化——模型自己参与优化自己的训练与推理系统，自主优化推理系统使端到端吞吐较基线提升 31.8%。需要谨慎读：这是「初步的」闭环，不是完全自主的自我改进，模型参与的是研发流程的某些环节，而非整个研发本身。但方向值得注意：当模型能力足够强，「模型参与研发模型」就不再是科幻，而是工程现实。

发布节奏同样延续了腾讯的打法：产品矩阵同步首发。WorkBuddy/CodeBuddy（国内版及国际版）、元宝、ima 同一天接入，提供限时两周免费体验；API 定价输入 6 元、输出 18 元每百万 token，国际价 0.834/2.501 美元。腾讯没有单独为模型办一场发布会，而是让模型直接进入自家产品——「模型 + 产品矩阵」的组合，与阿里「千问办公」同场发布的打法如出一辙。

Hy4 preview 的意义不在单个数字。它把腾讯混元推进到国产开源旗舰的第一梯队，用 1M 上下文与生产力定位回应规模竞赛；「模型参与自身研发」的递归自我改进叙事，为开源模型的能力来源提供了新的想象空间。对开发者，这是一个 Apache 2.0、1M 上下文、49B 激活的开源生产力模型，可自由商用；对行业，国产开源旗舰的竞争从模型层延伸到「模型参与研发」的自我改进叙事。preview 先行、正式版跟进——Hy4 系列的下一批模型已经在路上。

On August 28, 2026, Tencent Hunyuan released and open-sourced Hy4 preview. A 770B-total, 49B-active MoE with context past 1M under Apache 2.0—these numbers moved Hunyuan from the domestic second tier into the first. But the most readable line in the release materials is not a parameter; it is: "preliminary recursive self-improvement loop."

On specs: Hy4 preview is a major step up from Hy3 (295B/21B, 256K context): 770B total (780B on Hugging Face including the built-in MTP speculative decoding layer), 49B active, context past 1M. In an internal blind test (163 experts, 203 engineering tasks) it scored 2.99/4.00, slightly above GLM-5.3 (2.92) and Kimi K3 (2.94)—read with the vendor-self-reported caveat, but the direction is clear: Hunyuan entered front-line competition among domestic open flagships. Self-reported DeepSWE is 64.3 (Hy3 was 28.0) and Terminal Bench 2.1 is 85.4, positioned "for productivity" across software engineering, office analytics, game development, and research.

"Recursive self-improvement" is the most distinctive narrative of this launch. Tencent says Hy4 preview participated in its own full R&D pipeline for the first time: training methods, data strategy, evaluation systems, and automatic operator optimization—the model helping optimize its own training and inference systems, with self-optimized inference raising end-to-end throughput 31.8% over baseline. Read carefully: this is a "preliminary" loop, not fully autonomous self-improvement; the model participates in parts of the R&D process, not the whole of it. But the direction deserves attention: when models are strong enough, "a model participating in building models" stops being science fiction and becomes engineering reality.

The launch rhythm continues Tencent's playbook: the product matrix debuted together. WorkBuddy/CodeBuddy (domestic and international), Yuanbao, and ima plugged in the same day with a two-week free trial; API pricing is ¥6 input and ¥18 output per million tokens (international $0.834/$2.501). Tencent did not hold a separate launch event for the model—it put the model directly into its own products. The "model plus product matrix" combo mirrors Alibaba's Qianwen Office same-stage launch.

Hy4 preview's meaning is not in any single number. It moved Tencent Hunyuan into the first tier of domestic open flagships, answering the scale race with a 1M context and productivity positioning; the "model in its own R&D" recursive self-improvement narrative opened a new imagination for where open-model capability comes from. For developers, this is an Apache 2.0 productivity model with 1M context and 49B active parameters, freely commercial; for the industry, domestic open-flagship competition extended from the model layer to the "model in its own R&D" narrative. Preview first, GA to follow—the next batch of Hy4 models is already on the way.