Thinking Machines’ Inkling Small: Less parameters, better scores

Thinking Machines, the research lab led by former OpenAI CTO Mira Murati has unveiled Inkling Small, a 276 billion parameter, 12 billion active language model. Despite having fewer active parameters than its predecessor, Inkling Small has shown superior performance on major benchmarks such as GPQA Diamond (89% vs 87%) and Humanity’s Last Exam (32% vs 30%).

The model uses 24,000 tokens on average per task, compared to 45,000 tokens for DeepSeek Flash. You can fine tune the model using Tinker Playground, which is accessible from your web browser, or download its weights from Hugging Face, where they’re available under Apache 2.0 license.

Once again, we’re reminded that there are multiple ways to make AI systems smarter without throwing massive amounts of compute at the problem. A more efficient architecture can rival the performance of behemoths while requiring much less compute — shifting the financial incentives for developers who are billed per token.