Inkling: Our Open-Weights Model

Inkling：我们的开放权重模型

Our mission is to build AI that extends human will and judgment. We have developed a platform that lets anyone customize models, previewed an AI system built for interactive collaboration, and published novel research. Today we are advancing our mission by releasing a model we trained from scratch with the full weights available, so that people can make it their own. 我们的使命是构建能够延伸人类意志与判断力的 AI。我们开发了一个允许任何人定制模型的平台，预览了一个专为交互式协作而构建的 AI 系统，并发表了多项创新研究。今天，我们通过发布一个从零开始训练且完全开放权重的模型，进一步推进了我们的使命，让人们能够打造属于自己的模型。

Our model, called Inkling, is a Mixture-of-Experts transformer with 975B total parameters, 41B active. It supports a context window of up to 1M tokens. It was pretrained on 45 trillion tokens of text, images, audio and video. It is the first in a family of models of different sizes: alongside it we are sharing a preview of Inkling-Small, a lighter-weight model with 12B active parameters, trained with a similar recipe, that achieves strong performance with even lower cost and latency. 我们的模型名为 Inkling，是一个拥有 9750 亿总参数（410 亿激活参数）的混合专家（MoE）Transformer 模型。它支持高达 100 万 token 的上下文窗口，并使用 45 万亿 token 的文本、图像、音频和视频数据进行了预训练。这是我们不同规模模型系列中的首款产品：与此同时，我们还分享了 Inkling-Small 的预览版，这是一个采用类似方案训练、拥有 120 亿激活参数的轻量级模型，它以更低的成本和延迟实现了强大的性能。

Inkling reasons natively over text, images, and audio, and balances cost with performance through efficient and controllable thinking effort. We trained it to be a broad, balanced foundation model: strong across many domains, flexible enough to adapt. Inkling is not the strongest overall model available today, open or closed. Instead, a combination of qualities makes it a good open-weights base for customization: multimodal capabilities, efficient thinking, and availability on Tinker for fine-tuning. Inkling 能够对文本、图像和音频进行原生推理，并通过高效且可控的思维过程平衡了成本与性能。我们将它训练为一个广泛且均衡的基础模型：在多个领域表现出色，且具备足够的灵活性以进行适配。Inkling 并非当今市面上最强的模型（无论是开源还是闭源），但其多模态能力、高效的思维方式以及在 Tinker 平台上进行微调的便利性，使其成为一个理想的定制化开放权重基础模型。

Inkling is just the start: our first release in a model family we will continue to build on. We want to make customization accessible for more use cases, so Inkling is available for fine-tuning on Tinker today. Picking the right base model to fine-tune is a qualitative judgment that combines measurable benchmarks with the unique feel of a model that comes from playing with it. To enable the latter we’re adding the Inkling Playground in the Tinker console: a developer-facing interface for chatting with Inkling. Inkling 仅仅是一个开始：这是我们将持续构建的模型系列中的首个版本。我们希望让定制化能够服务于更多应用场景，因此 Inkling 即日起可在 Tinker 上进行微调。选择合适的微调基础模型是一种定性判断，它既需要参考可衡量的基准测试，也需要通过实际使用来感受模型的独特体验。为了实现后者，我们在 Tinker 控制台中添加了 Inkling Playground：一个面向开发者的 Inkling 对话界面。

To show what customization means in practice, we asked Inkling to fine-tune itself. Using Tinker, the model wrote its own fine-tuning job, ran it, and evaluated the result. 为了展示定制化的实际意义，我们让 Inkling 对自己进行了微调。通过使用 Tinker，该模型编写了自己的微调任务、运行任务并评估了结果。

(Note: The following section contains the technical logs of the self-fine-tuning process, demonstrating the model’s ability to generate code, execute training, and update its own weights to achieve a specific constraint—in this case, a lipogram task avoiding the letter “e”.) （注：以下部分展示了自我微调过程的技术日志，演示了模型生成代码、执行训练并更新自身权重以实现特定约束的能力——在此案例中，任务是编写不包含字母“e”的唇形文字。）

Summary of the Self-Fine-Tuning Process:

自我微调过程总结：

• Story 1: Landing and first prompt (Initial interaction) 故事 1：登录与首次提示（初始交互）
• Story 2: Base model answer (Standard response) 故事 2：基础模型回答（标准响应）
• Story 3: Fine-tuning intent (Defining the goal: a lipogram model) 故事 3：微调意图（定义目标：唇形文字模型）
• Story 4: Rubric (Setting the “no ‘e’” constraint) 故事 4：评分标准（设定“不含字母 e”的约束）
• Story 5: Training (Executing the Tinker API job) 故事 5：训练（执行 Tinker API 任务）
• Story 6: Self-update (Loading new weights) 故事 6：自我更新（加载新权重）
• Story 7: Updated model answer (The model successfully responds without using “e”) 故事 7：更新后的模型回答（模型成功在不使用“e”的情况下进行回答）