Inkling Is Mira Murati’s First Model. Here’s How to Actually Use It Inkling 是 Mira Murati 的第一个模型。这里是实际使用它的方法

TL;DR

• Thinking Machines Lab, founded by OpenAI’s former CTO, launched Inkling, a 975-billion-parameter open-weights mixture-of-experts transformer with a 1M-token context window, designed for customization rather than raw performance.
• The model is intentionally not positioned as the strongest available, reflecting a strategic bet that future AI value lies in owned, adaptable models rather than rented APIs.
• Inkling supports native multimodal reasoning (text, images, audio), features a controllable thinking effort dial, and uses novel architectural choices like learned relative-position embeddings and local attention in most decoder layers.
• Open weights are published on Hugging Face, but training data and pipeline remain closed—distinguishing open weights from full open source.
• The model is paired with Tinker, a fine-tuning platform, emphasizing specialization over general-purpose capability.

Why It Matters

This launch signals a pivotal shift in AI deployment strategy: from relying on proprietary, API-based models to adopting open, customizable foundation models that organizations can own and adapt. For practitioners and enterprises, this suggests a growing preference for models that can be fine-tuned on proprietary data to outperform general-purpose models on specific tasks—offering better cost-efficiency, privacy, and alignment with domain needs. It also challenges the industry’s benchmark-driven narrative, prioritizing adaptability and ownership over peak performance.

Technical Details

• Architecture: Mixture-of-experts transformer with 975B total parameters, 41B active per forward pass; 66 decoder layers, most using local attention over small windows.
• Context & Training: 1M-token context window; pretrained on 45T tokens across text, images, audio, and video—enabling native multimodal reasoning without adapters.
• Positioning Scheme: Uses learned, input-dependent relative-position embeddings instead of rotary position embeddings, claimed to improve extrapolation to long sequences.
• Inference Control: Features a “thinking effort” dial allowing per-request trade-offs between computational cost and reasoning depth.
• Open Weights: Fully open weights on Hugging Face; however, training data and pipeline are not open—clarifying the distinction between open weights and open source.
• Sibling Model: Inkling-Small (12B active parameters) previewed but not fully released at launch.

Industry Insight

• The rise of open-weights models like Inkling reflects a growing enterprise demand for AI assets that can be owned, audited, and fine-tuned—especially as raw model access becomes commoditized and interchangeable.
• Companies should evaluate foundation models not just on benchmark scores but on their adaptability, multimodal breadth, and support for fine-tuning ecosystems (e.g., Tinker), as these will determine long-term value in specialized applications.
• The “not the strongest” positioning is a strategic differentiator: it signals a focus on customization and ownership, which may become a key competitive advantage for organizations seeking to build proprietary AI capabilities rather than depend on third-party APIs.

TL;DR

• Thinking Machines Lab发布首个模型Inkling，参数量达9750亿，激活参数4100亿，上下文窗口支持百万token，采用混合专家架构。
• Inkling强调“开放权重”而非“开源”，允许用户下载、修改和部署模型，但不提供训练数据和管道，定位为可定制的基座模型而非通用最强模型。
• 公司策略押注“拥有模型”而非“租用API”，认为未来价值在于企业基于自身数据微调模型，而非直接使用通用大模型服务。
• 技术细节包括可控推理成本调节机制、输入依赖相对位置编码、局部注意力机制等，旨在平衡性能与效率。
• 伴随推出小型版本Inkling-Small（120亿激活参数），并配套微调平台Tinker，形成完整生态闭环。

为什么值得看

这篇文章揭示了AI行业从“模型即服务”向“模型即资产”的战略转向，对关注模型部署、定制化需求和长期技术自主权的企业与开发者具有重要参考价值。Thinking Machines Lab以诚实定位挑战行业惯例，提供了一种不同于追逐榜单的竞争路径，值得从业者深入思考模型商业化与落地场景的匹配方式。

技术解析

Inkling采用混合专家（Mixture-of-Experts）Transformer架构，总参数量9750亿，每次前向传播仅激活4100亿参数，兼顾容量与计算效率。其上下文窗口支持百万token，适用于超长序列任务，如法律文档、代码库或视频分析。模型在45万亿token的多模态数据（文本、图像、音频、视频）上预训练，原生支持跨模态推理，无需额外适配器。

架构上，Inkling摒弃标准的旋转位置编码（RoPE），改用输入依赖的相对位置编码，声称在长序列外推表现更优；66层解码器中多数采用局部注意力机制，减少全局计算开销。此外，模型引入“思考强度”调节器，允许用户按需切换推理深度与成本，提升灵活性与实用性。

权重完全开放于Hugging Face，但明确区分“开放权重”与“开源”——用户可自由使用与修改模型，却无法复现训练过程。配套平台Tinker提供微调支持，鼓励企业基于自身数据构建专属模型，形成“基座+定制”的产品组合。

行业启示

随着通用大模型能力趋同且API成本下降，企业差异化竞争将转向“垂直适配能力”。Thinking Machines的策略表明，未来核心竞争力可能不在于模型本身的绝对性能，而在于是否提供易于定制、可本地部署、支持私有化训练的开放基座。

该模式推动AI基础设施从“黑盒服务”转向“可编程资产”，促使组织重新评估数据主权、模型所有权与长期维护成本之间的关系。对于开发者而言，掌握微调、量化、蒸馏等适配技能将成为关键竞争力，而不仅是调用API。

此趋势也可能加速开源生态中“基座模型+垂直工具链”的分层结构形成，类似Linux与Android的关系——底层开放可控，上层应用百花齐放。