Thinking Machines bets on open weights with Inkling

What this is about

Thinking Machines Lab introduced its first in-house model on July 15, 2026: Inkling. The model is available with open weights, handles text, images, and audio, and is positioned mainly as a base for customization.

That matters because Thinking Machines is not simply shipping another closed chatbot product. Mira Murati’s team is framing Inkling as an alternative to dependence on a handful of centralized model providers: teams with enough infrastructure can download the weights, fine-tune them, and integrate the model into their own products.

What Inkling actually does

Inkling is a multimodal mixture-of-experts model. According to the model card, it has 975 billion total parameters, with about 41 billion active parameters per processing step. It supports a context window of up to one million tokens and accepts text, image, and audio input.

The open weights are available through Hugging Face. At the same time, Thinking Machines offers Tinker, a service developers can use to fine-tune the model. The practical barrier remains high: the model card says the BF16 version needs at least 2 TB of aggregate GPU memory, while the quantized NVFP4 version lowers that to at least 600 GB.

Why it matters

Open-weight models shift power. They do not automatically make every company independent, but they give research teams, platform vendors, and large enterprises more control over deployment, privacy, cost, and specialization.

For developers, Inkling is less interesting as a ChatGPT replacement than as a component. A company could fine-tune it for code review, document analysis, voice workflows, or multimodal support cases. The limit is also clear: this is not a laptop model. Running it yourself requires serious infrastructure.

In plain language

Inkling is like a very large professional kitchen. The recipes are on the table, but not everyone has the space, staff, and equipment to cook them. For small teams the recipe is useful; for large teams it can become their own kitchen.

A practical example

A medical-device manufacturer with 120,000 internal maintenance reports could fine-tune Inkling through Tinker on its own error codes, component photos, and audio from service calls. Instead of spending 40 minutes searching per case, a technician might see three likely causes in 8 minutes. That would only be trustworthy if the answers were checked against real repair data.

Scope and limits

• Open weights do not automatically mean low costs; the hardware footprint is still large.
• Benchmark claims mostly come from involved or model-adjacent sources and need validation in production.
• Multimodality expands the attack surface: images, audio, and long contexts can create new failure and misuse paths.

SEO & GEO keywords

Thinking Machines Lab, Inkling, Mira Murati, open weights, Hugging Face, Tinker, multimodal AI model, mixture of experts, AI model release, developer AI