Murati's 975B Model Fine-Tuned Itself on Launch Day — and Cracked the Top US Open-Weight Spot

2026-07-20 · Source: Towards AI - Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Murati's Thinking Machines Lab launched its new 975B-parameter model, Inkling, with a unique self-fine-tuning demonstration. The model autonomously wrote and executed its own fine-tuning job via the company's Tinker API, evaluated the results, and loaded new weights to achieve a specific lipogram behavior, avoiding the letter "e" in its responses within seven steps. Following its July 15 release, Inkling achieved an Intelligence Index score of 41 from Artificial Analysis, positioning it as the top-scoring open-weight model from a US lab. This score surpasses NVIDIA's Nemotron 3 Ultra (38), Gemma 4 31B (29), and gpt-oss-120b (24). Notably, despite 18 months of development and a \$2 billion seed round, the lab openly stated that "Inkling is not the strongest overall model available today, open or closed."

Key takeaway

For Machine Learning Engineers evaluating advanced model capabilities, Inkling's demonstrated self-fine-tuning ability signals a critical shift towards autonomous model adaptation. You should investigate platforms offering similar API-driven fine-tuning to streamline deployment and reduce manual iteration cycles. While its 41 Intelligence Index score is impressive for an open-weight model, consider the lab's candid assessment that it is not the "strongest overall" when making strategic model selections.

Key insights

Inkling, a 975B-parameter model, demonstrated autonomous self-fine-tuning to adapt its behavior post-training.

Principles

• Models can autonomously generate and execute their own fine-tuning processes.
• Transparency about model limitations can be a strategic launch approach.

Method

Inkling's self-fine-tuning involved writing a fine-tuning job, executing it via the Tinker API, evaluating the outcome, and loading new weights into its coding harness.

In practice

• Using Tinker's playground for model interaction and testing.
• Applying self-fine-tuning for specific behavioral adaptations like lipograms.

Topics

• Inkling Model
• Self-Fine-Tuning
• Large Language Models
• Model Benchmarks
• Open-Weight AI
• Tinker API

Best for: Research Scientist, AI Engineer, NLP Engineer, AI Scientist, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• Complete Guide to Thinking Machines Inkling — Inkling is an open-weights multimodal MoE model prioritizing customization and domain-specific fine-tuning over benchmark supremacy.
• I Fine-Tuned One Model 3 Ways: The $50,000 Run Forgot More Than the $1,500 One — Expensive full fine-tuning can degrade a model's general ability more than cheaper parameter-efficient methods.
• 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.
• The Blank Slate AI Strategy — A "blank slate" strategy commercializes open-source AI by offering a free base model and charging for customization services.
• The Best Open 32B Model (Open Thoughts Agents) — Optimizing supervised fine-tuning data through a structured pipeline significantly enhances open-source AI agent performance and generalization.
• 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.

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 Towards AI - Medium.