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. 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. 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 co
核心贡献 · Key contributions
发布了 Inkling，一个开放权重的专家混合 Transformer，总参数 975B，激活参数 41B，支持高达 1M 的上下文窗口。 Releases Inkling, an open-weights Mixture-of-Experts transformer with 975B total and 41B active parameters, supporting up to 1M context tokens.
在 45T 个文本、图像、音频和视频 token 上进行预训练，实现了跨这些模态的原生多模态推理。 Pretrained on 45T tokens of text, images, audio, and video, enabling native multimodal reasoning across these modalities.
引入了可控的思考努力，允许用户在性能与 token 效率之间权衡，以更少的 token 达到相当的分数。 Introduces controllable thinking effort, allowing users to trade off performance against token efficiency, achieving comparable scores with fewer tokens.
展示了自我微调：Inkling 在 Tinker 上编写并运行自己的微调任务，更新权重以成为无字母 e 的模型。 Demonstrates self-fine-tuning: Inkling writes and runs its own fine-tuning job on Tinker, updating its weights to become a lipogram model.
针对校准的不确定性、指令遵循和反审查进行训练，提高了可信度和安全性。 Trains for calibrated uncertainty, instruction following, and censorship non-compliance, improving trustworthiness and safety.
将强化学习扩展到超过 3000 万次 rollout，推理性能呈对数线性提升，并涌现出简洁的思维链。 Scales RL to over 30M rollouts, achieving log-linear improvement in reasoning performance and emergent concise chain-of-thought.
局限 · Limitations
Inkling 并非整体最强的模型；它优先考虑广度和可定制性，而非峰值基准性能。 Inkling is not the strongest overall model; it prioritizes breadth and customizability over peak benchmark performance.
自我微调演示仅限于简单的无字母 e 任务；复杂的现实世界微调可能需要更多监督。 The self-fine-tuning demonstration is limited to a simple lipogram task; complex real-world fine-tuning may require more oversight.
安全评估是初步的；在 Tinker 上微调可能会改变安全行为，长期影响尚未完全评估。 Safety evaluations are preliminary; fine-tuning on Tinker may alter safety behavior, and long-term impacts are not fully assessed.
Inkling-Small 仅是预览版；其完整权重尚未发布，性能可能与最终版本不同。 Inkling-Small is only a preview; its full weights are not yet released, and its performance may differ from final version.
该模型的多模态能力是从头训练的，但在某些领域可能落后于专业模型。 The model's multimodal capabilities are trained from scratch but may lag behind specialist models in some domains.
论文章节 · Sections（共 20）
概述Overview
能力Capabilities
通用模型Generalist model
智能体式编码与工具使用Agentic coding and tool use
单次生成带嵌入式浏览器使用的 Web 应用One-shot web app with embedded browser use
风格统一的工件Cohesively styled artifacts
通过长循环精炼创建多人游戏Multiplayer game created through long refinement loop
环境与验证约束（请先阅读）Environment and verification constraints (read first)