继 4 月底 Hy3 预览版发布后,我们收集了来自 50 多个产品的反馈,并使用更高质量的数据扩大了后训练规模。今天,我们推出 Hy3,其性能超越同尺寸模型,并能与参数规模大 2-5 倍的主流开源模型相抗衡。同时,它在各类产品和生产力任务中也展现出显著的实用性提升。
更强的智能体能力
在 Hy3 预览版的基础上,我们进一步提升了后训练数据的质量和多样性,同时扩大了强化学习训练的规模。Hy3 在推理、智能体和长上下文任务上均取得了扎实的进步,与规模大得多的旗舰模型相比也毫不逊色。
在编程、办公、金融建模、前端设计和游戏开发等生产力场景中,Hy3 取得了显著进展,现已成为一个可靠且高性价比的模型选择。我们邀请了 270 位专家,使用他们工作中的实际任务进行了一次盲评,Hy3 得分为 2.67/4,优于 GLM-5.1 的 2.51/4。其优势在前端开发、数据与存储以及 CI/CD 任务中最为突出。
更可靠的产品体验
模型的有用性并不能完全通过基准测试来体现。基于广泛的产品反馈,我们识别并修复了以下问题,并持续收到产品团队的积极反馈:
工具调用与输出格式的稳定性:我们修复了多个基础可靠性问题,使模型在各类工具配置和输出约束下达到生产级标准。工具调用的错误恢复能力和整体效率均得到提升。Hy3 还能在不同智能体框架间进行泛化。在 SWE-Bench Verified 上,使用 CodeBuddy、Cline 和 KiloCode 等不同框架时的准确率差异保持在 4% 以内。
知识与反幻觉:遵循“有依据时作答,缺乏证据时说明,不混淆来源或编造数据”的原则,我们实施了细粒度的数据清洗和训练约束。在基于真实场景的内部评估中,Hy3 的幻觉率从 12.5% 降至 5.4%,常识性错误率从 25.4% 降至 12.7%。这些改进切实减少了事实混淆、编造和逻辑矛盾。
复杂上下文保留与多轮意图追踪:通过 SFT 和 RL 的联合优化,Hy3 在指代消解、省略恢复和多轮约束继承等操作痛点方面取得了改进。在内部综合多轮测试中,问题率从 17.4% 降至 7.9%。Hy3 在 MRCR 等长对话评估中也取得了显著提升。其输出更加简洁,同时确保复杂意图在长程交互中不会衰减或偏移。
在 WorkBuddy 的内部测试中,我们看到任务成功率从 Hy3 预览版的 72% 提升至 Hy3 的 90%,平均完成时间下降了 34%。改进是广泛的——数据处理、文档工作和研究报告分析都明显变好。模型 token 效率也表现突出。我们在几个常见任务上将 Hy3 与 GLM-5.2 进行了对比:Hy3 在文档处理上少用了 47.4% 的 token,在演示文稿制作上少用了 49%。当模型能更正确地完成更多任务,且每个任务消耗更少 token 时,它自然就成为用户的日常首选。
Hy3 已在 GitHub、HuggingFace、ModelScope 和 AtomGit 上以 Apache 2.0 许可证开源。通过软硬件协同优化,Hy3 进一步降低了 API 价格,为更多用户和使用场景降低了采用门槛。
Hy3 API 定价(每 1M tokens)
| 输入 | 输出 | 缓存输入 | | --- | --- | --- | | 1 元 | 4 元 | 0.25 元 |
从 1 月底重建基础设施,到 4 月发布 Hy3 预览版,再到今天推出 Hy3 并伴随多种产品部署,我们在六个月内完成了一个端到端的模型开发循环。混元的重建与演进才刚刚开始。我们深知仍有许多挑战。我们将继续扎实地扩大训练规模、提升数据质量并优化用户体验细节,同时保持敏捷、透明和开放。
附录:基准测试结果
Following the Hy3 preview launch in late April, we gathered feedback from 50+ products and scaled up post-training with higher quality data. Today, we introduce Hy3, which outperforms similar-size models and rivals flagship open-source models with 2-5x parameters. It also shows significant gains in utility across various products and productivity tasks.
Stronger Agent Capabilities
Building on Hy3 preview, we further improved the quality and diversity of post-training data while scaling up RL training. Hy3 shows solid gains across reasoning, agentic, and long-context tasks, competitive with much larger flagship models.
In productivity scenarios such as coding, office work, financial modeling, frontend design, and game development, Hy3 has made remarkable progress and can now serve as a reliable, cost-effective model option. We ran a blind evaluation with 270 experts using tasks from their work, and Hy3 scored 2.67/4, outperforming GLM-5.1 at 2.51/4. The advantage was most substantial in frontend development, data & storage, and CI/CD tasks.
More Reliable Product Experiences
Model usefulness is not fully captured by benchmarks. Based on extensive product feedback, we identified and fixed the following issues, receiving consistently positive feedback from product teams:
Stability of tool calls and output formats: We fixed multiple baseline reliability issues, bringing the model to production-grade standards across tool configurations and output constraints. Tool-call error recovery and overall efficiency improved. Hy3 also generalizes across different agent scaffoldings. On SWE-Bench Verified, accuracy variance across scaffoldings like CodeBuddy, Cline, and KiloCode remains within 4 percents.
Knowledge and anti-hallucination: Guided by the ideal of "answer when grounded, state when evidence is missing, do not conflate sources or fabricate data," we implemented fine-grained data cleaning and training constraints. In internal evaluations based on real-world scenarios, Hy3's hallucination rate dropped from 12.5% to 5.4%, and commonsense error rates fell from 25.4% to 12.7%. These improvements materially reduce fact conflation, fabrication, and logical contradiction.
Complex context retention and multi-turn intent tracking: Through joint optimization of SFT and RL, Hy3 improved on operational pain points like coreference resolution, ellipsis recovery, and multi-turn constraint inheritance. On internal comprehensive multi-turn tests, the issue rate dropped from 17.4% to 7.9%. Hy3 also improved markedly on long-dialogue evals like MRCR. Its outputs are more concise while ensuring complex intents do not decay or drift over long-horizon interactions.
In our internal tests at WorkBuddy, we saw task success rate rise from 72% with Hy3 preview to 90% with Hy3, and average completion time dropped by 34%. The improvement was broad — data processing, document work, and research-report analysis all got noticeably better. Token efficiency stood out, too. We compared Hy3 with GLM-5.2 across a few common tasks: Hy3 used 47.4% fewer tokens for document processing, and 49% fewer for presentation creation. When the model gets more tasks right and uses fewer tokens per task, it naturally becomes users’ daily driver.
Hy3 is open-sourced under the Apache 2.0 license on GitHub, HuggingFace, ModelScope, and AtomGit. Through hardware-software co-optimizations, Hy3 further reduced the API price, lowering the barrier to adoption for more users and use cases.
Hy3 API Pricing(per 1M tokens)
| Input | Output | Cached Input | | --- | --- | --- | | 1 RMB | 4 RMB | 0.25 RMB |
From rebuilding infrastructure in late January, to releasing Hy3 preview in April and Hy3 today with various product deployments, we went through an end-to-end model development loop within six months. The rebuild and evolution of Hunyuan is only beginning. We are keenly aware that many challenges remain. We will continue to solidly scale up training, improve data quality, and optimize user-experience details, while staying agile, transparent, and open.
Appendix: Benchmark Results