Inkling 2 Small weights are public

Thinking Machines Lab's post on Inkling 2 Small, updated 12 November 2026, says the weights are now available for download on Hugging Face. The same note says the model is integrated with Unsloth, SGLang, vLLM, and TokenSpeed, can be post-trained with Miles, SkyRL, and Snowflake Cortex Training, and is available by API at NVIDIA, Fireworks, Together, Baseten, Nebius, Modal, and Cloudflare.

The model card (opens in a new tab) lists the license as Apache 2.0. It describes a mixture-of-experts model with 594 billion total parameters and 17 billion active. The main NVFP4 checkpoint needs a GPU cluster with at least 500 GB of aggregated VRAM, for example 2 NVIDIA B300 GPUs or 4 H200 GPUs. The MXFP8 checkpoint needs at least 800 GB, for example 4 B300 GPUs or 8 H200 GPUs. The lab recommends the NVFP4 checkpoint. It says MXFP8 scores 1 to 2 percentage points higher on evals, and that it does not consider that gap meaningful in practice.

The 5 November note on this site said the weights were still closed, with an Apache 2.0 release intended after safety testing. The weights are downloadable now.

Who it's for: a team that can host a checkpoint of this size, or call the model through one of the named APIs.

Fits: Best for Vibe Coding stays Cursor. A downloadable model is not the editor. Best for Writing stays Claude. The coding guide is unchanged. The verdict is unchanged.