NEWS → · Open weights · Thinking Machines
Inkling 2 Small weights are out: the 594B model lands on Hugging Face under Apache 2.0
Nov 12, 2026
Thinking Machines kept its promise on schedule: Inkling 2 Small, launched closed on its Tinker platform on Nov 5, 2026, put its full weights on Hugging Face on Nov 12, 2026 under Apache 2.0. It is a 594-billion-parameter mixture-of-experts model with 17 billion active per token and image input — and serving it yourself starts with a multi-GPU cluster: the model's own card puts the native NVFP4 checkpoint at a GPU cluster with at least 500 GB of aggregated memory, met by 2× B300 or 4× H200.
Image: Thinking Machines via thinkingmachines.ai
What shipped, from Thinking Machines' own Nov 12, 2026 update to its launch post and the model's Hugging Face card, both read first-hand: the weights are downloadable under Apache 2.0 (the card's licence field reads apache-2.0), integrated on day one with SGLang, vLLM, TokenSpeed and Unsloth, post-trainable with Miles, SkyRL and Snowflake Cortex Training, and available by API at NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare. The model itself: a 48-layer decoder-only MoE routing each token to 8 of 384 experts, a 292-million-parameter vision encoder for image input, variable reasoning effort, and a context window of up to 1 million tokens. At its Nov 5 launch the lab reported a score of 31 on the Artificial Analysis Intelligence Index — the lab's figure, not measured by us.
The hardware, in the card's own words: the main NVFP4 checkpoint — the precision the model was trained in, and the one the card recommends — requires a GPU cluster with at least 500 GB of aggregated VRAM, with 2× NVIDIA B300 or 4× NVIDIA H200 as the card's example configurations. The alternative MXFP8 checkpoint requires at least 800 GB, met by 4× B300 or 8× H200, and the card says its 1–2-point evaluation gain over NVFP4 is not meaningful in practice. There is no single-card route to this model at any precision the maker ships.
What it means if you rent GPUs
This is a cluster rental, priced like one. At the board's checked prices of Nov 12, 2026, the card's two NVFP4 example setups price out close together: 4× H200 at RunPod Community's $3.59/hr per GPU is $14.36/hr in total, and 2× B300 at Hyperstack's $7.40/hr is $14.80/hr — roughly $10,500–$10,800 a month run around the clock, before storage and your own serving work. That is the honest comparison to set against the hosted APIs the same announcement lists. The model is now in our catalogue (linked below), where its page works through every fit from the same memory arithmetic. A deployment guide — what it actually costs to run Inkling 2 Small well, and where the API beats renting — is proposed to the editor, not published by this sweep.