Inkling 2 Small
296 points
• 6 days ago
• Article
Link
Thinking Machines 推出了 Inkling 2 Small，这是其第二代 Mixture-of-Experts 模型，拥有 5940 亿个总参数和 170 亿个激活参数。该模型支持图像输入，并提供可变的计算投入水平。在 agentic coding 、 vision 、 knowledge work 和 reasoning 等任务上，相比前代产品展现出显著的性能提升。目前，该模型可在 Tinker 平台上用于推理与微调；在完成全面的外部安全测试后，计划以 Apache 2.0 许可证发布权重。
该模型在长周期的自主任务中表现尤为突出，例如通过程序化方式生成复杂环境，如城市模拟器或 PC 模拟器。其编码能力也通过在大规模数据集上进行强化学习得到增强，这些数据集引入了 agentic graders，使模型能更好地理解并执行指令，同时还能自主验证自己的输出。除软件开发外，Inkling 2 Small 在自动化研究（autoresearch）方面同样出色，已证明其能够在预测类任务中自主完成训练：通过反复构建评估脚本并持续打磨数据集，让模型逐步改进。
视觉理解是其另一项关键能力，配备了一个 2.92 亿参数的视觉编码器。模型在需要 OCR 、文档与信息图表解析，以及视觉推理的任务上表现非常出色。它能将视觉反馈有效嵌入工作流程，使其能够校验并优化自己生成的代码。此外，在知识密集型领域（包括金融、法律工作和定量分析）中，它也展现出强劲的能力，能够在阅读复杂文档与处理数据时做出接近专家水平的判断。
Inkling 2 Small 的研发以安全与对齐为核心。研究人员采用"纵深防御"的思路，通过同步与异步监控来识别并缓解不当行为。这包括对训练轨迹进行筛选，以抑制持续但不诚实的策略，并使用在策略内（on-policy）的自蒸馏，帮助模型内化更积极的行为准则。尽管在部分进攻性网络安全场景中，该模型仍未达到开放权重的前沿水平，但与既往版本相比，它在对齐基准测试上取得了明显进步。
Inkling 2 Small 的训练过程十分严苛：预训练阶段使用了超过 8 万亿个 tokens，并进行了为期九天的大规模联合强化学习，累计消耗超过 1700 万次 rollouts 。通过使用 NVIDIA GB300 机架以及 NVFP4 浮点格式等高效基础设施，团队能够对性能扩展实现可预测的提升。尽管如此，开发者也承认仍存在一些限制，例如缺少音频支持、交流偶尔偏啰嗦，以及在部分与安全相关的问题上存在过度敏感等情况，这些都将是后续优化的方向。
Thinking Machines has introduced Inkling 2 Small, their second-generation Mixture-of-Experts model featuring 594 billion total parameters and 17 billion active parameters. This model supports both image inputs and variable effort levels, showing significant performance gains over its predecessor across agentic coding, vision, knowledge work, and reasoning tasks. It is currently available for inference and fine-tuning on the Tinker platform, with plans to release the weights under an Apache 2.0 license following comprehensive external safety testing.
The model demonstrates notable strength in long-horizon autonomous tasks, such as procedurally generating complex environments like city simulators or PC emulators. Its coding capabilities have been enhanced through reinforcement learning on large-scale datasets that utilize agentic graders, allowing the model to better follow instructions and verify its work autonomously. In addition to software development, Inkling 2 Small excels at autoresearch, having proven its ability to autonomously train models on predictive tasks by iteratively building evaluation scripts and refining datasets.
Visual interpretation is another core capability, supported by a 292-million-parameter vision encoder. The model performs exceptionally well on tasks requiring OCR, document and infographic interpretation, and visual reasoning. It effectively integrates visual feedback into its workflows, enabling it to validate and refine its own generated code. The model also shows strong performance in knowledge-heavy domains, including finance, legal work, and quantitative analysis, where it demonstrates expert-level judgment when navigating complex documents and data.
Safety and alignment were central to the development of Inkling 2 Small. Researchers utilized a defense-in-depth approach, implementing both synchronous and asynchronous monitoring to detect and mitigate misbehavior. This included filtering training traces to discourage persistent but dishonest strategies and employing on-policy self-distillation to help the model internalize positive conduct principles. While the model remains behind the open-weight frontier in some offensive cyber scenarios, it shows a significant improvement in alignment benchmarks compared to previous versions.
The training process for Inkling 2 Small was rigorous, involving over 8 trillion tokens in the pre-training phase and a massive joint reinforcement learning run that consumed more than 17 million rollouts over nine days. By leveraging efficient infrastructures like NVIDIA GB300 racks and NVFP4 floating-point formats, the team was able to scale performance predictably. Despite its advancements, the developers acknowledge certain limitations, such as a lack of audio support, occasional verbosity in communication, and over-sensitivity in some safety-related queries, which remain areas for future refinement.
92 comments • Comments Link
- 近来开源权重模型的进展非常迅猛：在提供高质量的替代方案、足以取代既有最先进系统的同时，也体现出在模型架构和训练效率方面的显著提升。
- 行业内正出现一种"主权化"的开源 AI 趋势，背后推动力量来自那些把建立自身智能基础设施放在首位、同时又力图避免在地缘政治层面对境外提供商形成潜在依赖的企业与组织。
- 使用 NVFP4 精度进行训练已被证明是一项突破：相较于传统方法，它让规模较小的团队也能以惊人的速度和效率训练出能力很强的模型。
- 性能基准仍然是一个饱受审慎怀疑的话题，因为用户发现，报告中的数据往往无法真实反映现实中的效用；这促使更多人转向亲自上手测试，并对模型的"个性"与逻辑进行主观评估。
- 蒸馏依然是一种争议性但普遍的做法：一些团队公开使用来自更大模型的合成数据来指导训练；另一些则尝试从原始数据混合构建模型，以尽量减少对外部依赖。
- "品酒测试"看似有些琐碎，但作为一种有效的、非基准化的展示方式，它能体现模型底层的空间、视觉与推理能力，更强调泛化而非针对特定 benchmarks 的纯粹训练。
- 诸如硬件需求等现实约束仍是终端用户面临的重大障碍：高参数量（例如 596B）使得本地推理对除资源最丰富的少数环境外的大多数用户而言都成本过高。
- 速度与响应性正越来越被视为关键的"生活质量"指标：用户愿意在 agentic 设计系统中换取更快迭代，从而对推理能力做出一定程度的取舍。
- 关于前沿模型与开源权重替代方案之间存在 12 个月差距的看法仍然颇具争议，因为历史趋势表明，开源模型的追赶速度往往比许多专家预测得更快。
- 定价竞争力正在成为推动采用 AI 的主要因素：用户质疑，在像 DeepSeek 这类已有的低成本模型已经能为绝大多数专业任务提供足够效用的情况下，为边际收益支付额外溢价究竟是否值得。
这段对话体现了人们对"本土开源权重"快速发展的兴奋，也反映出在 AI 落地方面更务实、并注重成本的态度。尽管在技术讨论中 benchmarking 和理论能力占据主导，但实际共识更集中在"效用"上：用户主要想要的是能对编码与推理任务提供可靠、快速且具成本效益的自动化工具。关于国际模型与本土模型的争论，主要围绕信任与主权展开：人们明显更偏好透明度以及模型独立运行的能力，尽管普通开发者仍持续受到硬件条件的限制。 • Recent advancements in open-weight models demonstrate rapid progress, offering high-quality alternatives to established state-of-the-art systems while showcasing significant gains in model architecture and training efficiency.
• The industry is seeing a notable trend toward "sovereign" open-source AI, driven by corporations and organizations that prioritize owning their intelligence infrastructure while avoiding potential geopolitical dependencies on foreign providers.
• Training models using NVFP4 precision has proven to be a breakthrough, enabling smaller teams to train highly capable models with impressive speed and efficiency compared to traditional methods.
• Performance benchmarks remain a subject of healthy skepticism, as users observe that reported numbers often fail to capture real-world utility, prompting a preference for hands-on testing and subjective evaluation of model "personality" and logic.
• Distillation remains a controversial but prevalent practice, with some labs openly using synthetic data from larger models to bootstrap their training, while others attempt to build from original mixtures to avoid dependency.
• The "wine tasting" test, while seemingly frivolous, serves as an effective, non-benchmarked demonstration of a model's underlying spatial, visual, and reasoning capabilities, highlighting generalization over pure training for specific benchmarks.
• Practical constraints like hardware requirements remain a significant hurdle for end-users, as high parameter counts (e.g., 596B) make local inference prohibitively expensive for all but the most well-resourced environments.
• Speed and responsiveness are increasingly viewed as core quality-of-life metrics, with users willing to accept slight trade-offs in reasoning capabilities for the ability to iterate faster within agentic design systems.
• There is a recurring skepticism regarding the 12-month gap between frontier models and open-weight alternatives, as historical trends suggest that open models catch up much faster than many experts predict.
• Pricing competitiveness is becoming a primary factor for adoption, with users questioning the value proposition of paying premium costs for incremental gains when existing low-cost models like DeepSeek already provide sufficient utility for the vast majority of professional tasks.
The conversation reflects a blend of excitement for the rapid velocity of domestic open-weights development and a pragmatic, cost-conscious approach to AI adoption. While benchmarking and theoretical capabilities dominate the technical discourse, the practical consensus centers on utility: users are primarily looking for tools that offer reliable, fast, and cost-effective automation for coding and reasoning tasks. The debate over international vs. domestic models is largely framed around trust and sovereignty, with a clear preference for transparency and the ability to run models independently, despite the persistent hardware limitations faced by the average developer.