← All posts · News
Inkling 2 Small weights released under Apache 2.0
Thinking Machines' 594B MoE is now downloadable and live on seven API hosts.
Thinking Machines Lab released the weights of Inkling 2 Small on Hugging Face on November 12, 2026, under the Apache 2.0 license. The lab launched the model on November 5 with closed weights on Tinker and said the weights would follow after external safety testing (launch coverage).
What changed
- Weights are public as
thinkingmachines/Inkling-2-Small(NVFP4, the precision the model was trained in) andthinkingmachines/Inkling-2-Small-MXFP8, both under Apache 2.0. - The model is a 594B-parameter Mixture-of-Experts with 17B active, routing each token to 8 of 384 experts. It takes text and image input, returns text, and supports up to 1M tokens of context.
- Thinking Machines lists local serving recipes for SGLang, vLLM, TokenSpeed and Unsloth, and post-training support in Miles, SkyRL and Snowflake Cortex Training.
- API access is available through NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare, per Thinking Machines.
Third-party prices
Per million tokens, from each host's model page on November 13, 2026.
| Host | Model ID | Input | Cached input | Output |
|---|---|---|---|---|
| Together AI | thinkingmachines/Inkling-2-Small | $0.30 | $0.03 | $1.00 |
| Cloudflare Workers AI | @cf/thinkingmachines/inkling-2-small | $0.30 | $0.10 | $1.00 |
Running it yourself
Hardware guidance from the model card:
- NVFP4 checkpoint: at least 500GB of aggregated VRAM, for example 2x NVIDIA B300 or 4x NVIDIA H200.
- MXFP8 checkpoint: at least 800GB, for example 4x NVIDIA B300 or 8x NVIDIA H200.
Thinking Machines recommends the NVFP4 checkpoint for all use cases.
Inkling 2 Small is not in the RouterPlus catalog as of November 13, 2026. The catalog lists the earlier Inkling Small as thinkingmachines/inkling-small.
