Thinking Machines makes its first model a customization play
Thinking Machines Lab has released Inkling, its first in-house AI model, and the launch is less about topping every public leaderboard than about giving developers another high-end base model they can modify. WIRED reported the release on 15 July 2026, and Thinking Machines published its own technical announcement and model card the same day.
The company describes Inkling as an open-weights Mixture-of-Experts transformer with 975 billion total parameters and 41 billion active parameters per request. The model supports a context window of up to 1 million tokens and was pretrained on text, images, audio and video. Its output modality is text, including code and structured responses.
That combination puts Inkling in a specific part of the AI market: it is not presented by the company as the strongest model available, open or closed, but as a broad multimodal foundation model that can be adapted through fine-tuning.
What developers actually get
The practical release has three access paths. Thinking Machines says Inkling is available for fine-tuning through Tinker, its customization service. The model card says the weights are available through Hugging Face, and it also points to third-party inference providers for API access.
The open-weights claim matters because it lets researchers and companies inspect, host and fine-tune the model weights directly, subject to the release terms. It does not mean the training data or full source code are public. That distinction is important for teams evaluating transparency, compliance and reproducibility.
The model card also makes clear that self-hosting the largest checkpoint is not a casual workstation task. Thinking Machines says the BF16 checkpoint requires a GPU cluster with at least 2 TB of aggregate VRAM, while a quantized NVFP4 checkpoint lowers the requirement to at least 600 GB. Hugging Face's launch post similarly frames direct deployment as a large-infrastructure workload, while noting support in major inference engines such as Transformers, SGLang and vLLM.
For most developers, the immediate question is therefore not just whether Inkling is downloadable. It is whether Tinker, hosted inference providers, or smaller future variants make customization practical without buying a large GPU cluster.
Why the release is notable
Thinking Machines is led by former OpenAI CTO Mira Murati and has attracted attention because of its team and funding. But Inkling is the first major test of the company's product direction: decentralized, user-shaped AI rather than a single closed assistant experience.
The company says Inkling was trained as a generalist model across agentic, reasoning, coding, instruction-following, factuality, vision and audio tasks. It also previewed Inkling-Small, a lighter model with 12 billion active parameters, although the main release materials describe the smaller version as a preview rather than the main downloadable release.
One of the more interesting demonstrations in the company announcement is a self-fine-tuning workflow. Thinking Machines says it asked Inkling to use Tinker to create data, run a fine-tuning job, evaluate the result and switch to updated weights for a narrow behavior. That is a product demonstration, not independent proof of broad autonomous improvement, so it should be read as an example of the customization workflow the company wants to sell.
What to watch next
Inkling arrives as open-weight AI models are becoming a serious alternative to closed API-only systems for companies that want more control over data handling, cost and product behavior. The trade-off is operational complexity: running, securing, evaluating and fine-tuning a model at this scale requires infrastructure and expertise.
For AI builders, the useful signal will be adoption rather than launch-day benchmark tables. Watch whether developers can run Inkling reliably through common inference stacks, whether Tinker reduces the friction of fine-tuning, and whether independent evaluators confirm the model's strengths and limitations across coding, multimodal and long-context tasks.
The release gives Thinking Machines a concrete product around its decentralization message. Now the harder test is whether teams outside the lab can turn those open weights into useful, maintainable systems.