Thinking Machines发布了第二款重磅模型Inkling-Small[1]。这款模型总参数量为276B，激活参数仅12B，体量为初代Inkling的四分之一[1]。该模型采用MoE架构，具备原生多模态能力和100万token上下文[1]，在ARC-AGI-2等核心指标上刷新开源SOTA[1]，并在数学、推理、智能体编码、多模态基准中的性能堪比或超越初代Inkling和DeepSeek V4 Flash[1]。
Inkling-Small的训练采用了创新方法，先由初代Inkling进行on-policy蒸馏生成preview检查点，随后进行两周的智能体编码强化学习[1]。模型的硬件部署门槛明显降低，BF16检查点从2TB降至NVFP4量化版600GB起步，使中等规模团队也可进行模型定制[1]。
翁荔官宣离职仅两天后，OpenAI证实其已回归并带队攻坚「递归自我提升」(RSI)研究[1]。Thinking Machines六位联创中已有四人先后离职回归OpenAI，其中包括Barret Zoph、Luke Metz和翁荔[1]。
Thinking Machines has released Inkling-Small, its second major large language model, featuring 276 billion total parameters with 12 billion activated parameters [1]. The model employs a mixture-of-experts architecture, incorporates native multimodal capabilities, and supports a context window of one million tokens [1]. According to performance benchmarks, Inkling-Small achieves state-of-the-art results on open-source metrics including ARC-AGI-2 and matches or exceeds the capabilities of its larger predecessor Inkling and DeepSeek V4 Flash across mathematics, reasoning, agent coding, and multimodal tasks [1].
The model's development employed a two-stage training methodology: the team first conducted on-policy distillation using Inkling to generate preview checkpoints, followed by two weeks of reinforcement learning focused on agent coding [1]. Notably, the hardware requirements have been significantly reduced—the BF16 checkpoint has been scaled down from 2 terabytes to 600 gigabytes using NVFP4 quantization, enabling mid-sized teams to perform model customization [1].
The launch comes just two days after co-founder Weng Li announced his departure, with OpenAI subsequently confirming his return to lead efforts on "recursive self-improvement" (RSI) [1]. This development reflects organizational changes at Thinking Machines, where four of the company's six co-founders have now left to rejoin OpenAI, including Barret Zoph, Luke Metz, and Weng Li [1].
