Inkling AI: Mira Murati’s Open-Source Breakthrough 🧠

When Mira Murati departed OpenAI as Chief Technology Officer in late 2024, the artificial intelligence industry watched closely to see what her next move would be. Following her brief tenure as interim CEO during OpenAI’s high-profile board drama in late 2023, Murati struck out on her own, founding Thinking Machines Lab in February 2025.

After months of silent development—and securing a staggering $2 billion in seed funding at a $12 billion valuation—Thinking Machines Lab has officially launched its debut model: Inkling.

Inkling is not just another large language model; it is a massive, multimodal AI trained entirely from scratch, released with open weights under an Apache 2.0 licence. While it does not claim total dominance over the global market, Inkling provides Western enterprise teams and independent developers with something they have sorely lacked: a sovereign, highly capable, open-source alternative for complex agentic tasks.

What Is Inkling? Architecture and Technical Specifications

At its core, Inkling is built on a Mixture-of-Experts (MoE) architecture. Rather than activating the entire neural network for every single query, an MoE system selectively routes tasks to specialised sub-networks. This design keeps processing speeds fast and computationally efficient without compromising on context depth.

• Parameter Count: Inkling features 975 billion total parameters, with 41 billion active parameters per task. Due to its sheer size, running this model locally is out of reach for consumer hardware, but cloud deployment and fine-tuning are readily accessible.
• Multimodal Capabilities: Trained from scratch on 45 trillion tokens spanning text, images, audio, and video, Inkling processes and generates insights seamlessly across visual, vocal, and textual mediums.
• Context Window: The model boasts an expansive 1 million token context window, enabling it to analyse and reason over roughly 750,000 words in a single prompt sequence.
• Licencing & Accessibility: Fully available on Hugging Face, the weights can be downloaded without restrictions under the permissive Apache 2.0 licence. Developers can also fine-tune Inkling directly through Thinking Machines’ proprietary cloud environment, Tinker.

Benchmark Breakdown: Where Inkling Excels

Rather than hyper-focusing on niche benchmarks, Thinking Machines Lab designed Inkling as a versatile, well-rounded generalist. However, its standout capabilities emerge in autonomous agentic workflows—tasks where an AI assistant must reason, plan, and use external tools to complete multi-step software problems.

Real-World Agentic Tool Use
On the MCP Atlas benchmark, which measures how accurately an AI agent uses the Model Context Protocol to interact with external services, Inkling achieved a completion rate of 74.1%. This places it nearly 30 percentage points ahead of its primary Western open-weights competitor, Nvidia’s Nemotron 3 Ultra.

Software Bug Resolution
Evaluating an AI’s ability to fix actual GitHub bugs autonomously, the SWE-Bench Verified benchmark saw Inkling score 77.6%, once again outperforming Nemotron’s 70.7% mark.

Safety and Refusal Precision
Balancing strict safety protocols without triggering over-zealous refusals on benign prompts is a frequent challenge in open-weights models. On the FORTRESS Adversarial safety evaluation, Inkling scored 78.0%, the highest rating recorded among open-weights architectures in its category.

The Geopolitical Dimension: A Western Alternative to Asian Open-Source Leaders

To evaluate Inkling fairly, it is essential to look at the broader global landscape. Leading Chinese open-weights models continue to hold top spots across several raw performance metrics. For example, Z.ai’s GLM 5.2 leads on Terminal Bench 2.1 with an 82.7% success rate compared to Inkling’s 63.8%, while Moonshot AI’s Kimi K2.6 excels at advanced PhD-level scientific reasoning.

Thinking Machines Lab candidly acknowledges that Inkling is not the absolute top-performing model across every test suite worldwide. However, its value proposition lies in regulatory compliance, security, and data sovereignty.

Many organisations in Western countries face strict legal and operational restrictions that prevent them from routing sensitive workflows or enterprise data through Asian-hosted infrastructure or models developed in China. For these teams, Inkling delivers a robust, highly fine-tunable, and trustworthy open-source foundation that aligns with Western regulatory and operational standards.

What Lies Ahead for Thinking Machines Lab

Alongside the release of Inkling, Thinking Machines Lab previewed a lighter variant: Inkling-Small. Featuring 276 billion total parameters and 12 billion active parameters, this compact model already matches its larger sibling across several core reasoning benchmarks. Its open weights are slated for release as soon as initial testing concludes.

Despite early volatility in funding talks—with reported discussions around a $50 billion valuation stalling earlier in 2026—Thinking Machines Lab has proven its capacity to deliver frontier-class software. By placing open-weights tools into the hands of developers worldwide, Murati’s team is helping shape a more open, competitive, and accessible AI ecosystem.

To read the original report and explore additional details surrounding this launch, visit the full news article at Decrypt:

👉 Mira Murati Drops Her First AI Model After Leaving OpenAI—And It's Fully Open Source

Disclaimer: This article is provided for informational purposes only, mistakes may be made, and it's not offered or intended to be used as legal, tax, investment, financial, or any other advice.