Inkling
Thinking Machines Lab · Vision-language · MoE 975B (A41B) · 1000k context · Released 15 July 2026
Inkling was released on 15 July 2026 by Thinking Machines Lab, and it is a sparse mixture-of-experts with 256 routed experts and 2 shared, 6 routed active per token, across a 66-layer decoder with a hybrid of local and global attention reaching a 1-million-token context. It takes text, images and audio as input and produces text, including code and artifacts. A distinctive feature is controllable thinking effort, a dial from roughly 0.2 to 0.99 that trades answer quality against token cost. It is server-class, but the community has produced a broad range of self-quantised GGUF builds, including a smaller Inkling-Small variant, so quantised local deployment is more available than the headline size suggests.
Strengths
- Apache 2.0, fully permissive with no revenue gate or territory restriction
- Multimodal input across text, images and audio
- Controllable reasoning effort, so token cost can be traded against quality
- Very sparse, with only about 41 billion of 975 billion parameters active per token
Weaknesses
- Server-class at 975 billion parameters; even low-bit quants are large
- The smallest coherent quantisations carry a real quality trade-off
- Independent benchmark picture is still forming
- Image and audio support in local runtimes varies by tool
Hardware requirements
| Quantisation | Approx. VRAM | Notes |
|---|---|---|
| MXFP4 (community 4-bit) | ~490GB | Community self-quantised build with expert FFNs in MXFP4 and attention, router and shared experts at higher precision; roughly 490GB, multi-GPU server territory. Figure is approximate. |
| IQ1_M (smallest coherent) | ~200GB | The smallest community build that stays coherent, roughly 200GB, with a heavy quality trade-off. The separate Inkling-Small variant is more accessible again. Figures are approximate. |
What you'd need to run this
Licence
Apache 2.0 — read the licence
Inkling: common questions
- What hardware do I need to run Inkling?
- At its most compressed (IQ1_M (smallest coherent)) it needs roughly 200GB of VRAM. VRAM figures are approximate and depend on context length and settings.
- Is Inkling free for commercial use?
- Yes. Inkling is licensed under Apache 2.0, which permits commercial use with no meaningful conditions.
- Can I run Inkling on Apple Silicon?
- It can run on Apple Silicon through general runtimes, but it is not specifically optimised for it.
- What is Inkling's context window?
- Inkling has a context window of 1,000,000 tokens, about 1000k.
Availability
Recommended for
- Teams wanting a fully permissive, frontier-scale multimodal open model
- Server-class self-hosting where reasoning-effort cost control matters
- Long-context work over mixed text, image and audio inputs
Related models
Related guides
Catalogue entry last verified 24 September 2026. Specifications change; verify anything you are about to spend money on.