Thinking Machines Lab, the AI research company founded by former OpenAI CTO Mira Murati, has released Inkling, an open-weight model with 975 billion total parameters and 41 billion active parameters. The Mixture-of-Experts transformer supports a 1-million token context window and was pretrained on 45 trillion tokens of text, images, audio, and video, according to the company’s announcement.
Inkling is available on Hugging Face with full weights and can be fine-tuned through Thinking Machines’ Tinker platform. Fast Company reports that the company positions Inkling as a philosophical alternative to Anthropic’s proprietary model strategy, betting that enterprise and developer communities value locally-adaptable, agent-ready models over pure capability leadership.
Technical Specifications
Inkling is a Mixture-of-Experts architecture with 975B total parameters, of which 41B are active per forward pass. The model supports text, image, and audio inputs natively, with a context window of up to 1M tokens (64K/256K on the Tinker platform). Controllable thinking effort lets users adjust reasoning depth to balance latency against performance, a feature particularly relevant for agent workloads where cost-per-task varies by complexity.
Alongside Inkling, Thinking Machines previewed Inkling-Small, a lighter model with 12B active parameters trained with a similar recipe, targeting lower-cost and lower-latency deployments.
The company explicitly states that Inkling “is not the strongest overall model available today, open or closed.” Instead, it emphasizes breadth, multimodal capabilities, and suitability as a customization base. Benchmark comparisons published on the Inkling product page show competitive performance across general intelligence, agentic coding, audio/image input, and instruction following against Nemotron 3 Ultra, GLM 5.2, GPT 5.6 Sol, and Claude Fable 5.
The Customization Thesis
Thinking Machines’ core argument: the most valuable model for any specific agent deployment is one fine-tuned to that deployment’s domain, not the highest-scoring frontier model on general benchmarks. Inkling ships with fine-tuning support through Tinker, including a demonstration where the model wrote its own fine-tuning job, ran it, and evaluated the result.
This positions Inkling in a growing market segment alongside Chinese open-weight vendors. Moonshot’s Kimi K3 (2.8 trillion parameters, open-weight, launching July 27), DeepSeek V4-Pro, and GLM 5.2 all target builders who want to deploy customizable models rather than rent capability from closed APIs.
Two Markets Forming
The agent infrastructure market is splitting into two tiers. Closed frontier models, including Anthropic’s Claude Mythos and OpenAI’s GPT-5.6, serve regulated enterprises and high-security deployments where vendor trust and compliance matter more than per-token cost. Open-weight models like Inkling, Kimi K3, and DeepSeek V4-Pro target agent builders who need local deployment, domain-specific fine-tuning, and cost control.
Thinking Machines’ team, drawn from OpenAI, Character.ai, Mistral, and Meta’s AI research division (contributors to PyTorch, Segment Anything, and Fairseq), represents a notable concentration of frontier-model experience now directed at the open-weight side of that split.