# Thinking Machines unveils Inkling 2 Small with large agentic gains and higher per-task cost

published: 2026-11-06T16:05:35.453Z
topic: infrastructure
source: https://the-decoder.com/thinking-machines-inkling-2-small-claims-top-spot-among-western-open-weight-models/
url: https://aiagentnews.io/stories/thinking-machines-unveils-inkling-2-small-with-large-agentic-gains-and-higher

Thinking Machines has unveiled Inkling 2 Small, a 594-billion-parameter mixture-of-experts model that scores 31 on the Artificial Analysis Intelligence Index. Its weights are due under an Apache 2.0 license after external safety testing, but for now the model is available only through the company's Tinker platform.

Thinking Machines, the startup led by former OpenAI CTO Mira Murati, says Inkling 2 Small is the strongest Western open-weight model. The mixture-of-experts model has 594 billion parameters, with 17 billion active at any given time, and accepts text and images as input. Its adjustable thinking depth lets users trade speed for thoroughness.

On the Artificial Analysis Intelligence Index the model scores 31, 18 points above its predecessor Inkling Small and roughly level with Google's Gemini 3.8 Flash. The biggest gains are in agentic tasks, with AutomationBench-AA rising from 5 to 46 percent. The model produces about four times as many output tokens, and at $1.18 per task it costs nearly six times as much as its predecessor, though still less than Gemini 3.8 Flash at $7.77 or Claude Sonnet 5.5 at $10.02.

Thinking Machines says the model sometimes communicates inefficiently and does not always pick the best tools for large codebases. It does not support audio, is less optimized for chat, and can overreact to health-related topics. The company also plans to release more variants in the model family soon.

## Why it matters if you build agents

Developers evaluating open-weight models for agent workloads have a new Western option to test, but its higher token use means cost per task may rise compared with the predecessor. Because the weights are not yet released, self-hosting and fine-tuning are not possible today, and access runs through Tinker. Measure cost per task on your own workloads before switching.