Thinking Machines Lab Unveils Inkling, Its First Open-Weights Multimodal AI Model

Thinking Machines Lab has released Inkling, its first general-purpose artificial intelligence model, giving developers access to a nearly one-trillion-parameter system that can process text, images, and audio. The July 15, 2026 launch represents the most significant product milestone yet for the heavily funded AI company founded by former OpenAI Chief Technology Officer Mira Murati.

Rather than claiming to have built the world’s most capable model, Thinking Machines is positioning Inkling as a customizable foundation for companies, researchers, and developers that want greater control over how their AI behaves. Its weights can be downloaded through Hugging Face under an Apache 2.0 license, while the model can also be fine-tuned through Thinking Machines’ Tinker training platform.

That positioning could prove important as the AI market divides between proprietary frontier systems controlled by a small number of companies and open-weight alternatives that organizations can modify, host, and integrate into their own infrastructure.

A 975 Billion-Parameter Model That Uses Only a Fraction at Once

Inkling contains 975 billion total parameters, making it one of the largest open-weight models released to date. However, it uses a sparse Mixture-of-Experts architecture, meaning only 41 billion parameters are activated while processing any individual token.

The model contains 66 transformer layers and 256 routed experts. Each token is sent to six of those experts, along with two shared experts that remain active across requests. This approach is intended to provide the knowledge capacity of a very large model without requiring every parameter to be used during every inference operation.

Thinking Machines said the Mixture-of-Experts design largely follows ideas introduced in DeepSeek-V3, including its routing and load-balancing methods. Inkling also uses a combination of sliding-window and global attention layers, along with relative positional embeddings rather than the Rotary Positional Embeddings commonly used in modern language models.

The model supports a context window of up to one million tokens. Through Tinker, developers currently have access to 64,000-token and 256,000-token configurations, while the full context capacity is available through compatible deployments of the downloaded weights.

Native Reasoning Across Text, Images, and Audio

Inkling is natively multimodal, accepting text, images, and audio as inputs while generating text as its output. Images are processed using a hierarchical patch encoder, while audio is converted into discrete tokens and placed into the same shared representation space as text and visual information.

Thinking Machines pretrained the model on 45 trillion tokens drawn from text, images, audio, and video. The company said the data included publicly available material, licensed or acquired third-party sources, and synthetically generated or augmented content. However, the model card does not provide a complete dataset inventory, making “open weights” a more accurate description than fully open source.

The company also introduced a controllable reasoning system that allows developers to adjust how much computation Inkling applies before producing an answer. Lower reasoning settings can reduce latency and token consumption, while higher settings can be used for more difficult coding, mathematical, scientific, or tool-use tasks.

Alongside the main model, Thinking Machines previewed Inkling-Small, a 276-billion-parameter Mixture-of-Experts model with 12 billion active parameters. The company said improvements to its data and training recipe allowed the smaller model to match or exceed the larger version on several evaluations, although Inkling-Small has not yet received the same full release treatment.

Designed to Be Customized Through Tinker

Inkling is closely tied to Tinker, the cloud-based training platform that Thinking Machines introduced in 2025. Tinker allows developers to write training procedures locally while the platform manages the distributed GPU infrastructure required to fine-tune models ranging from relatively small systems to trillion-parameter Mixture-of-Experts architectures.

The platform supports supervised fine-tuning, reinforcement learning, Direct Preference Optimization, distillation, and Low-Rank Adaptation. Customers can download the resulting checkpoints, and Thinking Machines says customer training data is used only to fine-tune the customer’s models rather than being incorporated into its own foundation models.

To demonstrate this customization process, the company asked Inkling to fine-tune itself into a model that would never use the letter “e.” Inkling generated its own training data and evaluation method, created and ran a fine-tuning job through Tinker, tested the updated checkpoint, and then loaded the modified weights into its working environment.

The demonstration is deliberately narrow, but it illustrates the broader strategy. Thinking Machines wants developers to treat foundation models as starting points that can be adapted to an organization’s terminology, workflows, data, policies, and preferred behavior rather than as finished products controlled entirely by the original model provider.

Competitive Performance Without Claiming the Benchmark Crown

Thinking Machines openly acknowledges that Inkling is not the strongest model available across every category. Its benchmark results generally place it among the more capable open-weight systems, while the leading proprietary models from companies including Anthropic, Google, and OpenAI retain an overall advantage.

Inkling scored 97.1% on the 2026 American Invitational Mathematics Examination benchmark and 87.2% on GPQA Diamond, which evaluates graduate-level scientific reasoning. It recorded 77.6% on SWE-bench Verified and 54.3% on SWE-bench Pro Public, two benchmarks used to assess whether coding agents can resolve real software engineering problems.

On Humanity’s Last Exam, Inkling scored 29.7% without tools and 46% when given tool access. Its tool-enabled performance was competitive with several open models but remained below the strongest closed systems included in Thinking Machines’ comparison.

Inkling also received a score of 1,257 on Design Arena’s Agentic Web Development leaderboard, where human evaluators compare applications created by different models. That placed it close to several proprietary systems and ahead of a number of established open-weight competitors.

As with any vendor-published benchmark set, the results will require independent reproduction. Thinking Machines used external scores where available, but several evaluations relied on internal harnesses, configurations, or self-reported results from competing model providers.

Lower Costs Come With Significant Hardware Requirements

Inkling’s cost argument is based largely on its sparse architecture and controllable reasoning. Activating 41 billion parameters instead of all 975 billion should make inference considerably more efficient than running a similarly sized dense model, while allowing developers to disable extended reasoning for requests that do not require it.

Through Tinker, the 64,000-token version is listed at $1.87 per million prefill tokens, $4.68 per million sampled tokens, and $5.61 per million training tokens. The 256,000-token configuration is priced higher at $3.74 for prefill, $9.36 for sampling, and $11.23 for training, with lower rates available for cached input.

Those figures make large-model experimentation more accessible than assembling a dedicated training cluster, but Inkling is not a model that most developers will run on an ordinary workstation. The full BF16 checkpoint requires at least two terabytes of aggregated graphics memory, equivalent to eight Nvidia B300 GPUs or 16 H200 GPUs.

A quantized NVFP4 checkpoint reduces the requirement to approximately 600 gigabytes, but still requires infrastructure such as four B300 GPUs or eight H200 GPUs. In practice, many organizations are therefore likely to access Inkling through Tinker or third-party inference providers rather than directly hosting the complete model.

Balancing Resistance to Censorship With Safety Controls

One of Inkling’s more unusual design goals is what Thinking Machines describes as resistance to censorship. The company says it specifically trained the model to answer questions about topics that may be politically or institutionally sensitive rather than automatically refusing them.

Cognition evaluated Inkling using its Propaganda and Censorship benchmark, and Thinking Machines reported that the model demonstrated strong patterns of “censorship non-compliance.” The underlying idea is that an open model should not inherit unnecessary political restrictions or standardized opinions simply because those preferences were embedded by its original developer.

That does not mean Inkling is designed to answer every request. The company separately trained the model to refuse instructions related to weapons, violence, cyber abuse, chemical or biological threats, manipulation, and other dangerous activities.

Inkling scored 78% on the adversarial portion of the FORTRESS safety benchmark while correctly answering 95.9% of benign questions designed to resemble harmful requests. It also achieved 98.6% on StrongREJECT, which measures whether models refuse clearly harmful instructions. Thinking Machines says this combination indicates that the model can maintain safety controls without excessively blocking legitimate questions, although the effect of subsequent third-party fine-tuning remains an open research problem.

From OpenAI Leadership to a $12 Billion Startup

Thinking Machines Lab was formally launched in February 2025, several months after Murati left OpenAI. She had served as OpenAI’s Chief Technology Officer during the launches of products including ChatGPT, DALL-E, and the company’s early voice and multimodal systems.

The original team included approximately 30 researchers and engineers recruited from OpenAI, Meta, Mistral, and other AI organizations. Roughly two-thirds of its employees at launch had previously worked at OpenAI. The founding group included OpenAI co-founder John Schulman, Barret Zoph, Lilian Weng, Andrew Tulloch, and Luke Metz, bringing experience across reinforcement learning, model alignment, multimodal AI, pretraining, and post-training.

Only five months after its launch, Thinking Machines raised approximately $2 billion at a $12 billion valuation. Andreessen Horowitz led the round, with participation from Nvidia, Accel, ServiceNow (NOW ), Cisco, AMD, and quantitative trading firm Jane Street.

Nvidia has since expanded its relationship with the company through a multi-year strategic partnership. Announced in March 2026, the agreement calls for the deployment of at least one gigawatt of next-generation Nvidia Vera Rubin systems to support Thinking Machines’ model training and customizable AI platforms. Nvidia also made an additional investment in the company.

Inkling Reflects a Broader Bet on Decentralized AI

Thinking Machines argues that most AI systems are trained by a small number of companies and then distributed as largely fixed products. Its alternative is an ecosystem in which businesses and individuals can continually adapt models using their own knowledge, preferences, and judgment.

This philosophy helps explain why Inkling is being released with downloadable weights rather than offered exclusively through a proprietary interface. It also connects the model with Tinker, which provides the training infrastructure required to modify behavior without forcing every organization to assemble its own large GPU cluster.

There is still tension between this vision and the practical realities of the release. Inkling is open enough to download and modify, but sufficiently large that only well-funded organizations can realistically host the full model. Its training data is described in broad categories rather than fully documented, and customization can potentially weaken the safety controls built into the original checkpoint.

The release nevertheless gives Western developers another large open-weight option at a time when many of the most competitive customizable models have come from Chinese laboratories such as DeepSeek, Alibaba, Moonshot AI, and Zhipu AI. Reuters noted that Western open-weight development has lost momentum as several leading American companies have shifted toward more proprietary strategies.

What Inkling Means for AI

Inkling’s significance may depend less on whether it leads every benchmark and more on whether developers can turn it into specialized systems that outperform general-purpose models within narrower domains.

Companies in finance, healthcare, engineering, cybersecurity, scientific research, and software development often possess valuable internal knowledge that is poorly represented in public training data. A capable open-weight base model, combined with accessible reinforcement learning and fine-tuning infrastructure, could allow those organizations to encode more of that expertise directly into model behavior.

At the same time, the model will need to prove that its efficiency claims translate into lower total deployment costs, particularly once infrastructure, inference volume, fine-tuning, monitoring, and safety testing are considered. Independent researchers will also need to examine its benchmark performance, multilingual capabilities, political neutrality, hallucination rates, and resistance to safety degradation after customization.

Thinking Machines entered the market with an unusually experienced team, considerable investor backing, and infrastructure commitments normally associated with much older AI laboratories. Inkling is the company’s first opportunity to demonstrate that those resources can produce more than ambitious research plans and training platforms.

The model does not displace the leading proprietary systems, andThinking Machines Lab does not claim that it does. Instead, Inkling offers a different proposition: that the future of advanced AI may be shaped not only by whoever trains the most powerful general-purpose model, but also by who gives developers the strongest foundation for building models of their own.