Inkling: Our Open-Weights Model
1197 points
• 1 day ago
• Article
Link
Thinking Machines 发布了 Inkling，一个拥有 9750 亿参数的 Mixture-of-Experts Transformer 模型。作为通用基础模型，Inkling 设计为便于定制，支持最高 100 万个 token 的上下文窗口，并可原生在文本、图像和音频上进行推理。此外，公司还提供了 Inkling-Small 的预览版，这是一个为降低延迟和成本而优化的 2760 亿参数版本。
Inkling 的一大特色是可控的推理开销，允许用户根据具体需求在性能与 token 效率之间进行权衡。其架构对多模态输入采用无编码器方法并使用相对位置嵌入，开发者发现这比常规方案在长序列外推方面更为有效。模型在 45 万亿个 token 上进行了预训练，并通过大规模强化学习进一步精炼；值得注意的是，训练过程中出现了链式思考推理随着时间自发压缩的现象。
该模型在设计上强调认识论，注重置信度校准和对指令的遵循。为提高可靠性，研究团队使用自动化评分系统训练 Inkling，其中包括通过网页搜索验证事实声明的 claims grader 。安全评估显示，Inkling 具备强大的内置防护能力，能够有效拒绝有害请求，并在良性任务上保持高性能。
为便于实际应用，Thinking Machines 通过其 Tinker 平台开放了 Inkling 的微调能力。该生态系统使用户能够定制模型、评估表现并将其部署到不同的基础设施提供商上。团队还演示了一个案例，Inkling 成功自我微调以采用特定风格约束，展示了其作为开发工作流中具代理能力伙伴的潜力。
Inkling 的完整权重已在 Hugging Face 上发布，其中包含为 NVIDIA Blackwell 系统优化的版本。通过提供基础模型与深度定制工具，Thinking Machines 致力于帮助组织构建反映专业知识和特定运营需求的 AI 。
Thinking Machines has released Inkling, a Mixture-of-Experts transformer model with 975 billion total parameters. Designed as a broad foundation model, Inkling is built to serve as a base for customization, supporting a context window of up to 1 million tokens and native reasoning across text, images, and audio. Alongside the primary model, the company has provided a preview of Inkling-Small, a 276-billion-parameter version optimized for lower latency and cost.
Inkling distinguishes itself through controllable thinking effort, allowing users to balance performance against token efficiency based on their specific needs. Its architecture utilizes an encoder-free approach for multimodal inputs and relative positional embeddings, which the developers found to be more effective for long-sequence extrapolation than standard alternatives. The model was pretrained on 45 trillion tokens and further refined through large-scale reinforcement learning, which, notably, resulted in an emergent compression of its chain-of-thought reasoning over time.
The model is designed with a focus on epistemics, emphasizing calibrated confidence and instruction following. To ensure reliability, researchers trained Inkling using automated grading systems, including a claims grader that verifies factual assertions via web search. Safety evaluations indicate that Inkling demonstrates strong built-in safeguards, effectively refusing harmful requests while maintaining high performance on benign tasks.
To facilitate practical application, Thinking Machines has made Inkling available for fine-tuning through their Tinker platform. This ecosystem enables users to customize the model, evaluate its performance, and deploy it across various infrastructure providers. Demonstrating the model's self-improvement capabilities, the team showcased an instance where Inkling successfully fine-tuned itself to adopt a specific stylistic constraint, illustrating its potential as an agentic partner in development workflows.
Inkling's full weights are available on Hugging Face, including versions optimized for NVIDIA Blackwell systems. By providing both the foundational model and the tools for deep customization, the company aims to support organizations in developing AI that reflects specialized knowledge and unique operational requirements.
286 comments • Comments Link
• Inkling 模型的发布标志着 American open-weight 研究迈出了重要一步：它提供了一个带音频能力的多模态、长上下文模型。
• 对基准测试应保持审慎，因为它们往往无法反映现实世界的效用或模型的"agentic"表现；在这些方面，一些体积更小或基准分数较低的模型常常优于那些被基准调优过的大型模型。
• Open-weight labs 的核心商业策略是构建无缝的 fine-tuning 和推理生态，试图通过 RLaaS（Reinforcement Learning as a Service）等服务来实现价值变现，而不仅仅依赖模型权重本身。
• 关于 AI 公司的"护城河"存在很大争议。许多人认为当前的融资模式更像是风险投资的慈善行为，因为多数 labs 都在靠消耗资本抢占市场份额，尚未验证出可持续的商业模式。
• 全球竞争是主要驱动力之一，尤其是各国希望保持 sovereign AI 能力，以规避可能由美国主导的对 American-hosted 模型施加的限制。
• 与 Tinker 等开发平台的集成表明，趋势正在推动可定制、可调节的模型成为企业更经济的选择，相比单纯依赖通用的大规模训练模型。
• 尽管长上下文能力常被吹捧，但很多从业者发现模型在超过 150k–200k tokens 后质量会下降，因此除非模型能持续保持严格的指令遵循，否则把上下文长度当成目标可能是一个陷阱。
• 人们持续担心模型评估的透明度，一些研究人员批评使用雷达图而不公开原始数据；同时也有声音欢迎这种具竞争力、可用的 open-weight 替代方案的出现。
• Thinking Machines 在 16 个月内取得的快速进展常被拿来与 Moonshot (Kimi) 等老牌参与者多年的努力相比较，这表明训练前沿级模型的门槛可能比此前预期的更低。
• 实际的 agentic 开发通常依赖于特定的、非基准的性能指标，例如调用工具的可靠性和指令遵循能力；这些指标在不同模型间差异巨大，比静态分数更能影响用户留存。
总体而言，讨论反映出对新兴 American open-weight AI labs 的谨慎乐观，但对其长期商业可行性仍持怀疑态度。社区对能否出现替代 GLM 和 DeepSeek 等主导参与者的新型、多模态且可定制的解决方案感到兴奋，但也对当前的投资是否可持续存在分歧。共识是：在现实世界中，agentic 工作流的实际效用目前比基准分数更重要，而这些模型的真正价值将取决于其周围开发者生态的质量以及解决特定企业任务的能力。 • The release of the Inkling model marks a significant step for American open-weight research, offering a multi-modal, long-context model that incorporates audio capabilities.
• Benchmarks should be viewed with skepticism, as they often fail to capture real-world utility or "agentic" performance, where some smaller or lower-scoring models outperform larger, "benchmaxxed" counterparts.
• A core business strategy for open-weight labs involves providing a seamless fine-tuning and inference ecosystem, aiming to capture value through services like RLaaS (Reinforcement Learning as a Service) rather than just the weights themselves.
• There is significant debate regarding the "moat" of AI companies, with many observers noting that current funding models resemble VC charity, as most labs are burning capital to gain market share without a proven, sustainable business model.
• Global competition is a major driver, particularly the need for nations to maintain sovereign AI capabilities to bypass potential US-led restrictions on American-hosted models.
• Integration with development platforms like "Tinker" suggests a push toward making bespoke, customizable models a more economical choice for enterprises than relying solely on generic, mass-trained models.
• While long-context performance is often touted, many practitioners find that model quality still degrades after 150k-200k tokens, making context length a potential "trap" unless the model maintains rigorous instruction following.
• Concerns persist about the transparency of model evaluations, with some researchers criticizing the use of radar plots over raw data, even as others appreciate the arrival of a competitive, usable open-weight alternative.
• The rapid progress of labs like Thinking Machines in 16 months is being compared to the years-long efforts of established players like Moonshot (Kimi), suggesting that the barriers to entry in training frontier-level models may be lower than previously assumed.
• Practical agentic development often hinges on specific, non-benchmark performance metrics like tool-calling reliability and instruction adherence, which vary wildly between models and influence user retention more than static scores.
The discussion reflects a cautious optimism regarding the emergence of new American open-weight AI labs, balanced against skepticism about the viability of their long-term business models. While there is enthusiasm for new, multi-modal, and customizable alternatives to dominant players like GLM and DeepSeek, the community remains divided on whether current investment levels represent a sustainable strategy or a speculative bubble. Ultimately, the consensus is that real-world utility in agentic workflows currently outweighs benchmark performance, and the true value of these models will be determined by the quality of the surrounding developer ecosystems and their ability to solve specific enterprise tasks.