昆仑万维推进天工 / Skywork 大模型

“天工”和“Skywork”长期并排出现，却承担着不同工作。天工是普通用户能打开的搜索与对话产品；Skywork 更像写给开发者和研究社区的署名。对昆仑万维而言，这种双品牌不是文字游戏。只有前端入口，容易被看成接入其他模型的壳；只有模型仓库，又很难解释能力怎样抵达真实用户。

Skywork 的基础模型项目在 2023 年进入开源社区，官方仓库称其预训练使用了 3.2TB 以中英文和代码为主的数据，并开放模型、部分数据、评测数据与评测方法。到 2024 年中，Skywork-MoE 等项目又把混合专家训练经验写进技术报告。单个榜单名次会变化，开放材料的作用却更持久：它让外界能够检查昆仑是否真的在训练和研究模型，而不只是购买 API 再包装。

另一端的天工把这种能力放进搜索增强对话和多模态产品。用户看到的是一个可以提问、查资料、生成内容的入口，不必理解底层仓库叫什么。产品端提供使用数据和分发，模型端提供可辨认的技术身份；两者一起构成昆仑在大模型竞争中的生存结构。它未必能在每条能力线上压过头部公司，却可以保留自己的模型名、用户关系和迭代节奏。

后来的发展证明这不是一条只做聊天的窄路。Skywork-o1 转向推理，SkyReels 把品牌延伸到以人物为中心的视频生成，更多多模态和 Agent 项目继续从同一组织下长出。不能把这些后续成果倒写成 2024 年中已经完成的矩阵；它们更像当时策略的结果——一旦保留研究团队、模型品牌和产品入口，新的分支就有地方附着。

Skywork 记录的是第二梯队公司的另一种选择。昆仑万维没有只押注一个爆款聊天应用，也没有把全部身份寄托在一次开源发布上，而是让天工与 Skywork 分别守住入口和基座。这个结构成本更高，路线也更杂，但它让公司在模型供应越来越集中的市场里，仍有机会决定自己的产品下一步接什么能力。

Tiangong and Skywork frequently appeared together, but they performed different jobs. Tiangong was the search-and-conversation product an ordinary user could open. Skywork was closer to a signature written for developers and the research community. For Kunlun, the dual branding was more than a naming exercise. A consumer interface by itself could be dismissed as a wrapper around somebody else’s API; a model repository by itself could not explain how the capability reached users.

The base-model Skywork project entered the open community in 2023. Its official repository described pretraining on 3.2 terabytes of data, primarily Chinese, English, and code, and released models alongside training data, evaluation data, and evaluation methods. By mid-2024, projects such as Skywork-MoE were documenting experience with mixture-of-experts training. Individual leaderboard positions would move. The open material had a more durable function: outsiders could inspect evidence that Kunlun was training and studying models rather than simply purchasing access and changing the label.

Tiangong placed that work at the other end of the chain in search-augmented conversation and multimodal products. A user encountered an interface for questions, research, and content generation without needing to know the repository name underneath. The product supplied distribution and usage feedback; the model organization supplied a recognizable technical identity. Together they formed Kunlun’s survival structure in the foundation-model market. The company did not need to lead every capability line to retain its own model name, customer relationship, and release cadence.

Later branches showed that the structure was not limited to chat. Skywork-o1 moved toward reasoning, SkyReels extended the organization into human-centered video generation, and additional multimodal and agent projects followed. Those later results should not be written backward as a finished matrix in mid-2024. They were consequences of the earlier choice. Once a research team, a model brand, and a product entrance remained in place, new branches had somewhere to attach.

Skywork records an alternative available to a second-tier internet company. Kunlun did not place its entire identity in one breakout chat application, nor did it rely on a single open release. Tiangong and Skywork separately guarded the entrance and the foundation. The structure was expensive and its roadmap could look untidy, but it left the company with some authority over which capability its products would absorb next in a market where model supply was becoming increasingly concentrated.