AI technology

Inkling 2 Mini brings open weights and a looped design to local hardware

Thinking Machines' November 20 release offers downloadable weights and Tinker access. Its looped architecture reuses middle layers, while the model card specifies memory requirements and the announcement acknowledges weaknesses in multi-turn chat.

By SpicyRanksPublished 2 linked sources ↓

Event date: Nov 20, 2026

A smaller model with repeated computation

Thinking Machines announced Inkling 2 Mini on November 20. Its looped design runs the middle layers three times per token. The company reports 26.4 billion total parameters and distinguishes 3.4 billion active parameters from 5.4 billion effective weight activations when repeated passes are counted. [1]

Spicy's takeAnalysis: reusing layers changes how much computation happens before another token appears. It does not make parameter count a complete measure of runtime cost. Compare the supported implementation and workload rather than assuming a smaller label means a particular speed on your machine. [1]

Read the hardware card before downloading

The model card describes text and image inputs with text output. It lists a mixed-precision checkpoint of about 20GB, support for one NVIDIA DGX Spark or Apple silicon devices with 32GB or more memory, and recommends at least 32GB VRAM for other GPU platforms. Tinker also offers hosted access. [2]

Spicy's takeAnalysis: decide whether you want a hosted experiment or a local deployment first. For local use, verify the exact runtime and memory configuration before budgeting hardware. This report has not installed the checkpoint, measured generation speed or tested the listed integrations. [2]

Capabilities come with limits

The announcement acknowledges that Mini is not its strongest small-model option for multi-turn chat or heavily constrained instructions, and trails larger models on some reasoning benchmarks. Its performance comparisons remain provider-reported here; no independent superiority claim is made. [1]

Spicy's takeAnalysis: the useful release milestone is access to a documented looped model with a concrete local deployment path. Whether it fits a companion interface, coding workflow or another application still requires task-specific evaluation. Downloadability is a starting point, not a finished product assessment. [1][2]

Sources & further reading

Source material checked Nov 22, 2026. Linked material may change after publication.

  1. Introducing Inkling 2 Mini ↗

    Thinking Machines Lab · Primary source · Nov 20, 2026

  2. Inkling 2 Mini model card ↗

    Thinking Machines Lab · Primary source