Mira Murati’s Thinking Machines Lab released Inkling, its first model trained from scratch. It is a mixture-of-experts model with 975 billion total parameters but activates 41 billion per token, lowering the work needed for each response. Inkling accepts text, images, and audio, can handle up to one million tokens of context, and lets developers control how much computation it spends reasoning. The full weights are downloadable under the Apache 2.0 license, subject to a separate Model Acceptable Use Policy, including a smaller NVFP4 version for Nvidia hardware. Thinking Machines also previewed Inkling-Small, with 276 billion total and 12 billion active parameters. The company says Inkling is not the strongest model overall. Its pitch is the combination of multimodal input, efficient reasoning, and weights that developers can alter.
The unusual offering is how the model connects to customization. Developers can fine-tune Inkling through Tinker, download the resulting weights, and run them elsewhere. Tinker limits Inkling training to 64K or 256K context windows. After a limited 50% discount, its 64K tier costs $1.87 per million input tokens, $4.68 per million generated tokens, and $5.61 per million training tokens. An Inkling Playground is also available in the Tinker console. In the company’s demo, Inkling wrote and ran its own fine-tuning job in about 27 minutes, then loaded the new checkpoint. That turns customization from a dedicated GPU project into an API job, although Tinker supports low-rank adaptation rather than changing every model weight.
Inkling separates access from practicality. Its estimated 1.9 TB BF16 checkpoint is genuinely downloadable, but most users will still need a cloud host. Open weights give companies control over deployment and tuning, while Tinker, hosted APIs, and hardware partners provide the expensive path from files to a usable product.
Sources:
- Thinking Machines Lab: Introducing Inkling
- Inkling model card and weights
- Tinker training API documentation
- Tinker models and pricing
Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service
Reuse
Citation
@misc{kabui2026,
author = {{Kabui, Charles}},
title = {Inkling: {Open} {Weights} {Built} for {Custom} {AI}},
date = {2026-07-19},
url = {https://toknow.ai/posts/inkling-open-weights-custom-ai/},
langid = {en-GB}
}