Thinking Machines、軽量モデル「Inkling-Small」正式公開 サイズ4分の1で「Inkling」に匹敵する性能

Summary

Thinking Machines Lab released a lightweight multimodal MoE Transformer called Inkling-Small, which is one-quarter the size of Inkling yet delivers comparable or superior performance across a range of benchmarks.

Why it matters

The model demonstrates that it is possible to reduce the number of parameters while preserving performance, potentially lowering entry barriers for researchers and businesses. Its open-weight version under the Apache 2.0 license encourages community-driven innovation, and its reduced hardware requirements make deployment more accessible.

Highlights

• Inkling-Small has a total of 2.76 trillion parameters but only 120 billion active MoE experts, and it outperforms Inkling on benchmarks like Humanity's Last Exam (31.6% vs 29.7%).
• Training enhancements—including policy distillation and two weeks of reinforcement learning for agent coding—contribute to the performance gains.
• The model's hardware footprint is dramatically smaller: a BF16 checkpoint runs on just 600GB of GPU memory compared to Inkling's 2TB, and the quantized version requires only 180GB.

Critical take

While Inkling-Small excels in speed and efficiency, the paper notes that Inkling still holds an advantage in knowledge coverage and factual accuracy; this highlights a trade-off between model size and comprehensive understanding.