Thinking Machines released Inkling 2 Small on November 5 with 594B total parameters and 17B active
Inkling 2 Small uses a sparse mixture-of-experts architecture with 594B total and 17B active parameters, supports image input and variable reasoning effort, scores 31 on the Artificial Analysis Intelligence Index, and weights are promised under Apache 2.0 after safety testing.
On November 5, Thinking Machines Lab released its second-generation model Inkling 2 Small. The model uses a sparse mixture-of-experts architecture with 594 billion total parameters and 17 billion activated per token, accepts text and image input and produces text output, does not support audio, and offers variable reasoning effort.
The model card states a maximum context window of 1 million tokens, while the Tinker platform currently serves 64K and 256K context variants. Evaluations ran at effort 0.99 and temperature 1.0. The Artificial Analysis Intelligence Index is given as 31 in the introduction and 30.6 in the detailed evaluation table, against 13 for the previous Inkling Small. Coding results include 86.3 percent on SWE-Bench Pro v2 Hard, 72.1 percent on DeepSWE and 25.4 percent on FrontierCode, while Terminal-Bench 4.0 came in at 13.1 percent, below GLM-5.3-Flash at 32.8 percent.
Weights were closed at launch, with inference and fine-tuning offered only through the Tinker platform, and the company said full weights would be released under an Apache 2.0 license the following week after external safety testing and red teaming. Pretraining ran 68 hours on NVIDIA GB300 NVL72 racks across stages totaling about 8.4 trillion tokens, followed by a 1075-step reinforcement learning run over nine days on 2048 GPUs.