Thinking Machines Lab releases Inkling 2 Small weights on Apache 2.0

Thinking Machines Lab published the open weights for Inkling 2 Small on Hugging Face at thinkingmachines/Inkling-2-Small under an Apache 2.0 licence on 12 November 2026, fulfilling the open-weight commitment made during its 5 November Tinker preview. Built as a general-purpose image-text-to-text model for agentic tool use, coding assistants and retrieval-augmented generation, the model accepts UTF-8 text and pixel images (optimised between 40 px and 4096 px per side) across a context window of up to 1M tokens.

Sparse MoE architecture, native NVFP4 weights and GPU footprints

Inkling 2 Small uses a 48-layer decoder-only transformer with hybrid local and global attention, a vision transformer encoder projecting images into a shared hidden space, and a sparse Mixture-of-Experts backbone routing each token to 8 of 384 experts (17B active parameters out of 594B total; Hugging Face lists 383B packed tensor parameters on the repo badge). Because pretraining, midtraining and post-training all ran natively in NVFP4, Thinking Machines Lab recommends the NVFP4 checkpoint, which fits in 500 GB of aggregate VRAM (such as 2 NVIDIA B300 GPUs or 4 H200 GPUs).

An alternative MXFP8 checkpoint requires at least 800 GB of VRAM (4 B300 or 8 H200 GPUs) for a 1 to 2 percentage point evaluation delta that the lab notes is negligible in practice. Local serving recipes ship for SGLang, vLLM, TokenSpeed, Unsloth, Hugging Face Transformers and Docker Model Runner, alongside fine-tuning via the Tinker Cookbook with tml-renderers or Snowflake Cortex Training.