Inkling 2 Small 发布：68 小时主预训练，多模态与可定制性并重
Thinking Machines 发布了 Inkling 2 Small ：一款支持文本、图像输入和文本输出的混合专家（MoE）模型，重点强调训练效率、原生视觉能力，以及面向具体任务的可定制性。 对开发者来说，首先要区分服务可用与权重开放：模型目前已在 Tinker 和 Tinker Playground 提供推理与微调服务，但 发布时尚未开放权重 。团队计划在外部安全测试完成后，于发布后的下一周以 Apache 2.0 许可证开放权重；这仍是计划，而非已经完成的发布动作。 核心规格与价格 项目 已披露信息 总参数量 5940 亿 激活参数量 170 亿 架构 多模态 MoE，具备原生视觉能力 输入与输出 文本、图像输入；文本输出 上下文窗口 100 万 token 最大输出长度 100 万 token 输入价格 每百万 token 0.30 美元 缓存输入价格 每百万 token…
Inkling 2 Small, a mixed expert (MoE) model developed by Thinking Machines, has been released with support for text, image input and text output. The model emphasizes training efficiency, native visual capabilities, and customization for specific tasks. Developers should note that the model is currently available for inference and fine-tuning through Tinker and Tinker Playground, but the weights have not yet been made publicly available.
The team plans to release the weights under an Apache 2.0 license following external security testing, which is scheduled for the following week.
Key specifications and pricing details are as follows:
- Total parameters: 5940 billion
- Active parameters: 170 billion
- Architecture: Multi-modal MoE with native visual capabilities
- Input and output: Text and image input; text output
- Context window: 1 million tokens
- Maximum output length: 1 million tokens
- Pricing: $0.30 per 100,000 tokens for input, $0.03 per 100,000 tokens for input caching, $1.00 per 100,000 tokens for output, and $1.50 per 100,000 tokens for training
The 68-hour figure represents the duration of the main pre-training stage on GB300, using 84 trillion tokens, and employing NVFP4 precision. The total training data size is estimated at around 1.2 quadrillion tokens, but this may include intermediate training stages and cannot be directly compared to the main pre-training stage. The post-training phase lasted for 9 days with approximately 17 million rollouts.
In the Artificial Analysis Intelligence Index v5.0.1, Inkling 2 Small scored 31 points, 18 points higher than Inkling Small. The score is comparable to Gemini 3.8 Flash (high) and higher than MiMo-V2.6-Pro, indicating an improvement in the model's performance on the overall evaluation. However, these results do not guarantee superior performance across all tasks. Additionally, there is a lack of comprehensive Muse Spark comparison data, so conclusions about the model's superiority should be avoided.
For engineering decision-making, the ranking score can serve as an initial filter, but practical task validation of Inkling 2 Small is still necessary. When discussing the model's customization capabilities, it is essential to consider both the basic capabilities and fine-tuning performance. The team has set customization evaluation tests as an important design goal, and the results indicate that Inkling 2 Small outperforms alternative models in fine-tuning under fixed data and budget conditions.
However, complete disclosure of tasks, comparison settings, and results is required to accurately assess the model's relative advantages in specific business scenarios. Both the team's internal reviews and the sharing of model selection experiences should mention the tasks, data volume, budget, and comparison settings to avoid misunderstanding the fine-tuning gains as universal benefits.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
