Inkling
Apache on the tin, terms in the drawer
What it is
Thinking Machines Lab's first open-weights model: a 975B-total / 41B-active sparse MoE (6 of 256 experts plus 2 shared) that takes text, images, and audio in, with a 1M-token context, trained on 45T tokens. Day-0 support landed in transformers, vLLM, SGLang, and llama.cpp, with an NVFP4 checkpoint alongside BF16.
Why it's interesting
The most-watched debut of 2026 — Mira Murati's lab shipping trillion-scale weights with frontier-class scores (77.6% SWE-Bench Verified, 46.0% HLE with tools). But the Apache-2.0 label comes with a separately binding Model Acceptable Use Policy, which is exactly the kind of rider this site exists to flag.
Use cases
- Self-hosted frontier-class assistants and agents
- Multimodal apps that need voice and vision in one model
- 1M-token document and codebase work
Who it's for
Well-resourced labs, inference providers, researchers
Setup
Advanced. ~2TB VRAM for BF16, ~600GB for the NVFP4 build; community 1-bit GGUFs squeeze it far smaller at a real quality cost
Limitations & cautions
The hardware appetite is enormous, the AUP restricts fields of use, and this is a v1 with no stability track record — the card itself warns against unvalidated high-stakes deployment.
Editorial takeaway
The year's biggest debut is also its best license quiz. Read past the tag before you build on it.