Thinking Machines Inkling: A Powerful Open-Weights Multimodal LLM

Thinking Machines Lab recently released Inkling, a standout open-weights model designed for reasoning, coding, and multimodal tasks. Released in mid-July 2026, Inkling features a massive 975B total parameters (41B active) Mixture-of-Experts (MoE) architecture with native support for text, image, and audio inputs, and a context window up to 1M tokens.

Key Features

• Variable Thinking Effort: Adjust reasoning depth to balance performance, speed, and cost — perfect for different use cases.
• Multimodal Native: Process images, audio, and text in a shared hidden space. Strong on vision (charts/diagrams) and audio understanding.
• Token Efficiency: Achieves competitive results using fewer tokens than many peers, making it more practical for long sessions.
• Customizable: Open weights enable fine-tuning on the Tinker platform for domain-specific needs.
• Strong Safety & Calibration: Excellent instruction following, calibrated confidence, and high safety scores.

Performance Highlights

Inkling scores impressively on key benchmarks:

• GPQA Diamond: 87.9%
• SWE-bench Verified: 77.6%
• Strong agentic coding and general reasoning, positioning it as a leading U.S. open-weights model.

In practical tests (e.g., via Command Code), Inkling excels at following complex constraints to generate functional scripts, such as system readiness checks for Google Agent Development Kit across multiple languages. It reasons through requirements accurately and produces working code without unnecessary installations.