Murati's Thinking Machines launches Inkling, 975B-param model leading US labs but trailing China

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has unveiled Inkling, a Mixture-of-Experts Transformer with 975 billion total parameters, of which 41 billion are active at any given time. Unlike many competing open-source models, Inkling natively supports text, images, and audio within a single model and offers a context window of up to one million tokens—substantially larger than many current alternatives. The weights are freely available on Hugging Face, and the company also provides access through Tinker, its platform for adapting AI models to specific tasks.

Thinking Machines pre-trained Inkling on 45 trillion tokens of public and synthetic text, images, audio recordings, and videos. The training dataset includes public data subject to intellectual property protection, and the company generated synthetic data using methods including the Chinese AI model Kimi K2.5, which also served as the basis for Cursor's coding model. The company explicitly acknowledges in its announcement that "Inkling is not the strongest overall model available today," positioning it instead as a flexible base model optimized for customization and efficiency.

On benchmarks, Inkling leads U.S. open-weights models but trails China's best. According to AI benchmarking platform Artificial Analysis, Inkling scores 41 on the Artificial Analysis Intelligence Index, making it the leading open-weights model from a U.S. lab—three points above Nemotron 3 Ultra (38) and well ahead of Gemma 4 31B (29) and gpt-oss-120b (24). On GDPval-AA v2, an agent-based benchmark simulating knowledge-work tasks, Inkling reaches an Elo rating of 1,238, beating Kimi K2.6 (1,190) and DeepSeek v4 Flash max (1,189). On the Tau-3 banking benchmark, Inkling scores 24 percent, ahead of DeepSeek v4 Flash max (23 percent) but behind Kimi K2.6 (21 percent). However, Inkling performs poorly on factual accuracy, scoring just +2 on Artificial Analysis's AA Omniscience benchmark with 40 percent accuracy and a 63 percent hallucination rate—results that are likely to limit its use in applications requiring highly accurate information.

Pricing for Inkling with a 64K context window is $1.87 per million input tokens and $4.68 per million output tokens; for context windows up to 256,000 tokens, pricing rises to $3.74 for input, $0.748 for cached input, and $9.36 for output. A key efficiency advantage is that Inkling averages just 25,000 output tokens per Intelligence Index task, substantially fewer than GLM-5.2 max (43,000), Kimi K2.6 (38,000), and DeepSeek v4 Pro max (37,000). Thinking Machines positions this efficiency through continuously adjustable "thinking effort," allowing users to choose their preferred balance between cost and performance while maintaining result quality.

Thinking Machines is also previewing Inkling-Small, a more compact variant with 276 billion total parameters and 12 billion active parameters. The smaller model delivers similar or better results on several benchmarks: it scores 88.3 percent on GPQA Diamond compared with 87.2 percent for Inkling, and on the HLE benchmark with tools it scores 46.6 percent versus Inkling's 46.0 percent. The company credits these improvements to changes in pre-training data and the training process, and plans to publish the full weights once testing is complete.

Thinking Machines Lab's release of Inkling marks Mira Murati's first major product launch since leaving OpenAI. The model reflects a strategic choice to compete on efficiency and task-specific adaptation rather than raw performance—a positioning that acknowledges the reality of the open-weights landscape. On the Artificial Analysis Intelligence Index, Inkling's score of 41 places it as the leading U.S. open-weights model, ahead of Nemotron 3 Ultra (38), signaling that Murati's team has succeeded in building a competitive system for American labs. However, the benchmarks also reveal the ceiling: Chinese models like Kimi K2.6 and DeepSeek v4 Flash max outperform Inkling on agent-based knowledge-work tasks (GDPval-AA v2), and Inkling's accuracy score of +2 on AA Omniscience with a 63 percent hallucination rate suggests it is less suitable for applications requiring high factual reliability.

The company's business model—positioning Inkling as a foundation for fine-tuning through the Tinker platform rather than as an all-purpose contender—mirrors a broader market shift toward customizable models. That approach is underscored by the model's multimodal support and million-token context window, both designed to enable flexible adaptation. The pricing structure, at $1.87 per million input tokens for a 64K window, sits slightly above comparable Chinese open-source models but benefits from Inkling's efficiency: it averages just 25,000 output tokens per Intelligence Index task, substantially fewer than GLM-5.2 max (43,000) or Kimi K2.6 (38,000). Thinking Machines' preview of Inkling-Small, which beats the larger model on some benchmarks like GPQA Diamond, further suggests the team is prioritizing practical utility and efficiency over headline performance metrics.

No comments yet. Be the first to share your thoughts!