Thinking Machines launches Inkling, a 975B open-source AI model
Thinking Machines has launched Inkling, a 975-billion-parameter open-weights AI model built as a customizable foundation for fine-tuning rather than a finished product. The Mixture-of-Experts system activates 41 billion parameters per task, reasons across text, images, and audio, and is available now on the company's Tinker platform and Hugging Face.
Key Takeaways
- Inkling is a 975B-parameter MoE model with 41B active parameters and up to 1M-token context.
- It was pretrained on 45 trillion tokens spanning text, images, audio, and video.
- Thinking Machines positions Inkling as a broad generalist base for customization, not a top leaderboard model.
- Fine-tuning is available today via Tinker, with full weights released on Hugging Face.
- A lighter Inkling-Small preview (12B active parameters) is also being shared.
Thinking Machines Lab released Inkling on July 15, 2026, marking its first model trained entirely in-house with full weights published for outside developers. The launch advances the startup's mission to build AI that extends human judgment by giving organizations a multimodal starting point they can adapt to their own workflows.
What Is Inkling and Why Does It Matter?
Inkling is a Mixture-of-Experts transformer with 975 billion total parameters, though only about 41 billion are active during any given task. That sparse design keeps inference faster and cheaper than a dense model of comparable scale would be.
The model supports a context window of up to one million tokens and was pretrained on 45 trillion tokens drawn from text, images, audio, and video. It reasons natively across those input types while producing text outputs, including code and structured data.
Thinking Machines explicitly says Inkling is not the strongest overall model available today, open or closed. Instead, the company highlights multimodal breadth, controllable thinking effort, and fine-tuning access as qualities that make it a practical open-weights base for real-world customization.
How Does Inkling Perform on Key Benchmarks?
Despite its generalist positioning, Inkling posts competitive scores on several widely watched tests. It reaches 77.6% on SWE-bench Verified for software engineering and 91.4% on VoiceBench for voice understanding, placing it among strong open-weights performers in those categories.
The model also supports adjustable thinking effort, letting developers trade token cost against reasoning depth. On Terminal Bench 2.1, Thinking Machines reports Inkling matching rival Nemotron 3 Ultra at roughly one-third the token spend.
For readers tracking the broader Future Tech & AI Wonders landscape, Inkling joins a crowded 2026 open-weights field where specialization and post-training flexibility matter as much as raw benchmark crowns.
How Can Developers Fine-Tune and Deploy Inkling?
Inkling is available for fine-tuning on Tinker today, with context-length options of 64K and 256K tokens. The company is offering a limited-time 50% discount on Tinker access and added an Inkling Playground in the Tinker console so developers can test the model before committing to a training run.
Full model weights are published on Hugging Face, as detailed in Thinking Machines' official announcement. An NVFP4 checkpoint is also available for efficient inference on NVIDIA Blackwell hardware. Hosted APIs are offered through partners including TogetherAI, Fireworks, Modal, Databricks, and Baseten.
Thinking Machines also released updated cookbooks and a tml-renderer tool to support post-training with tool calls, reasoning traces, and multimodal inputs. The lab says it is continuing to study how fine-tuning on Tinker affects safety behavior in customizable models.
What Comes Next for the Inkling Family?
Inkling is described as the first release in a growing model family. Alongside the flagship, Thinking Machines shared a preview of Inkling-Small, a lighter MoE variant with 12 billion active parameters that matches or exceeds its larger sibling on many benchmarks while cutting latency and cost.
The company says Inkling-Small weights will ship once final testing completes. Thinking Machines describes Inkling as just the start of a model family it plans to keep building on as customization becomes central to enterprise AI deployment.