❯ Thinking Machines releases Inkling 2 Small weights, bringing a 594 billion-parameter model trained natively in 4-bit precision

From hosted access to downloadable weightsThinking Machines Lab opened Inkling 2 Small weights, updating its announcement on November 12 with Apache 2.0 checkpoints on Hugging Face. The model debuted on November 5 through Tinker. It uses a mixture-of-experts architecture with 594 billion total parameters, about 17 billion active per processing step, and image-input support.

Low precision spans trainingPretraining, midtraining and post-training used NVFP4, the 4-bit floating-point format supported by Nvidia’s Blackwell architecture. The released low-precision checkpoint therefore matches its training precision. The model targets coding, visual understanding and long-running tool use, and can be customized through further training on Tinker.

Open weights still require substantial memoryThe official model card specifies at least 500GB of aggregate GPU memory for NVFP4 and 800GB for the alternative MXFP8 checkpoint. Multiple cloud API providers offer hosted access, giving developers a choice between operating it themselves and using a service. “Small” describes its place within the model family; deployment still targets multi-GPU servers.

▮ SIGNALNative low-precision training puts practical operating costs at the center of open-weight deployment, while the hundreds-of-gigabytes memory requirement keeps much of the deployment advantage with cluster operators and cloud providers.