Ex-OpenAI CTO Murati's Thinking Machines drops Inkling, a 975B parameter model that leads US labs but trails China

2026-07-16 · Source: The Decoder · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation, Robotics & Autonomous Systems · Depth: Intermediate, short

Summary

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has released Inkling, a 975 billion parameter multimodal open-weights model. This Mixture-of-Experts Transformer, with 41 billion active parameters, natively processes text, images, and audio, supporting a context window up to one million tokens. Artificial Analysis reports Inkling scores 41 on its Intelligence Index, making it the leading U.S. open-weights model, surpassing Nemotron 3 Ultra (38). It achieves an Elo rating of 1,238 on GDPval-AA v2, outperforming Kimi K2.6 (1,190) and DeepSeek v4 Flash max (1,189) on agentic tasks, and demonstrates high token efficiency, averaging 25,000 output tokens per task. However, Inkling exhibits a 63 percent hallucination rate and 40 percent accuracy on AA Omniscience, scoring only +2, and has higher costs (\$1.87/\$4.68 per million input/output tokens for 64K context) compared to some Chinese models. A smaller Inkling-Small (276 billion parameters) is also previewed, sometimes outperforming the larger model.

Key takeaway

For Machine Learning Engineers evaluating open-weights models for agentic applications, Inkling offers leading U.S. performance and token efficiency, especially for multimodal tasks. You should account for its 63 percent hallucination rate and higher cost compared to some Chinese alternatives. Consider fine-tuning Inkling for specific use cases where its agentic capabilities are critical, but implement robust factual verification layers. Explore Inkling-Small for potentially better performance on certain benchmarks.

Key insights

Inkling leads U.S. open-weights models in agentic tasks and efficiency but struggles with factual accuracy and cost.

Principles

• Multimodal support enhances base model flexibility.
• Agentic task performance can outweigh raw intelligence.
• Smaller models can sometimes outperform larger counterparts.

Method

Thinking Machines pre-trained Inkling on 45 trillion tokens, including public and synthetic data, using Kimi K2.5 for synthetic generation.

In practice

• Use Inkling as a base for fine-tuning specific multimodal tasks.
• Adjust "thinking effort" to balance cost and performance.
• Evaluate smaller models like Inkling-Small for specific benchmarks.

Topics

• Inkling
• Multimodal AI
• Open-weights Models
• Agentic AI
• Model Benchmarking
• Hallucination Rate
• Fine-tuning

Best for: CTO, VP of Engineering/Data, AI Architect, AI Scientist, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• Mira Murati’s Thinking Machines drops Inkling, an open-weights model anyone can access — Open-weight multimodal models like Inkling offer customizable, cost-effective alternatives to proprietary AI systems.
• Complete Guide to Thinking Machines Inkling — Inkling is an open-weights multimodal MoE model prioritizing customization and domain-specific fine-tuning over benchmark supremacy.
• 🗞️ Mira Murati's Thinking Machines Lab drops massive open-weight AI model, Inkling, Apache 2.0 license, no restrictions. — Agentic AI models and efficient inference techniques are rapidly advancing, driving both performance and cost-effectiveness.
• Kimi 3, Inkling: These Two Open Models May Change “Everything” — The release of Kimi K3 and Inkling, featuring massive scale and open weights, signals a potential paradigm shift in AI model utility.
• Murati's 975B Model Fine-Tuned Itself on Launch Day — and Cracked the Top US Open-Weight Spot — Inkling, a 975B-parameter model, demonstrated autonomous self-fine-tuning to adapt its behavior post-training.
• not much happened today — New AI models prioritize efficiency, multimodality, and reduced hallucinations, while some projects face leadership instability.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Decoder.