Thinking Machines Releases Inkling 2 Small Weights Under Apache 2.0, With Day-0 Support From SGLang, Snowflake and Cloudflare
The multimodal MoE model (594B total, 17B active, 1M-token context) was API-only at launch on 5 November; its Apache 2.0 checkpoints landed on Hugging Face on 12 November as promised.
Thinking Machines Lab has released the checkpoints of Inkling 2 Small on Hugging Face under the Apache 2.0 license, fulfilling the open-weights promise it made when the model launched API-only on 5 November. Two checkpoints are available, one a mixed NVFP4/MXFP8 release and one full MXFP8. The model is a mixture-of-experts network with 594 billion total parameters, 17 billion active per token, multimodal input and a 1 million token context window.
Ecosystem support arrived the same day. SGLang published day-0 serving benchmarks on GB300 and VR200 systems using DSpark with MegaMoE at TP4/EP4, Miles shipped reinforcement-learning post-training support, and Snowflake and Cloudflare added the model to Cortex Training and Workers AI respectively.
The release matters for two reasons. It is the first frontiers-class non-Chinese model under a fully permissive license that enterprises can self-host, an option for organizations that avoid both Chinese models and closed APIs. It also proves a new distribution pattern: closed API first to fund training and red-teaming, then open weights on schedule. Practical deployment remains demanding, with roughly 380 GB of memory needed at 4-bit precision, putting the model in multi-GPU rack territory rather than single-server setups.