A Year After Leaving OpenAI, Its Former CTO Unveils Her First In-House Model

Mira Murati served as OpenAI’s CTO before leaving in September 2024 and founding Thinking Machines Lab in 2025. The company recruited several researchers from OpenAI and focused on making models easier for businesses and developers to understand, modify, and train.

Thinking Machines Lab’s first product was Tinker, a model fine-tuning platform. Developers provide their data and scoring methods, and Tinker fine-tunes open-weight models. At the time, Thinking Machines Lab did not yet have a foundation model of its own.

On July 15, 2026, the company released Inkling, its first publicly available general-purpose large model trained from scratch. The full weights are on Hugging Face under the Apache 2.0 license, with direct Tinker integration. The company now has both its own model and its own training platform.

So What Exactly Is Inkling?

In short, Inkling is a multimodal foundation model that can be downloaded, deployed, and trained further. It accepts text, images, and audio, and can write code, use tools, and operate websites. It is not a standalone chat app for consumers. It is more like a general-purpose engine that developers can build into coding assistants, customer service systems, research tools, web agents, or internal enterprise software.

How Is It Different from a Typical Open Model?

Many open-weight releases focus on the weight files. Inkling was integrated with Tinker from day one. Developers can test the model in the Playground, send their own data, scoring rules, and training jobs to Tinker, produce a version better suited to their business, and then download it or deploy it through a third-party service.

What Can Inkling Actually Do?

Start with the overall picture. The official release compares Inkling with four representative models across ten evaluations. The thick blue line is Inkling: it does not dominate every rival on any single axis, but it has no obvious gaps across coding, tool use, vision, audio, factuality, and instruction following. That is exactly the profile of a general-purpose foundation that developers can continue training.

1. Scale: 975 Billion Parameters, Plus a Faster Small Version

Original Inkling has 975 billion total parameters and uses a mixture-of-experts architecture. It activates only about 41 billion parameters for each token and supports up to 1 million tokens of context. These are tokens, not characters; Tinker currently offers practical context tiers of 64K and 256K.

Inkling-Small preview has 276 billion total parameters and activates about 12 billion per token. With upgraded pretraining data and training methods, it approaches or even surpasses the original on many tests while running faster and costing less. It is not equally strong on every test, however, and the full Small weights have not been released.

2. Controllable Reasoning: Developers Choose the Depth

Many reasoning models decide for themselves how long to think. Even simple tasks can produce lengthy internal reasoning, increasing wait times and cost. Inkling gives developers that choice. The release page shows reasoning effort from 0.2 to 0.99: lower values produce faster answers and use fewer tokens, while higher values give the model more time for math, code, and multistep tasks.

3. Native Multimodality: Vision and Audio from the Start

Inkling’s vision and audio components were trained from scratch. Audio is converted into spectrograms the model can read, while images are split into 40×40-pixel patches. After lightweight conversion, they enter the same model alongside text; there is no need to transcribe everything into text first.

Audio: It can transcribe speech, follow spoken instructions, answer questions about recordings, and analyze longer audio.

Vision: It can interpret charts, formulas, and mathematical diagrams. When details matter, it can also use Python to zoom in or crop an image.

4. Getting Work Done: Less Guessing, More Follow-Through

Less guessing when uncertain: The training team designed a reward that allows the model to abstain. On short-answer questions it scores only for correct answers; when uncertain, saying “I don’t know” or qualifying its judgment is more rewarding. A separate grader searches the web to check facts in its answers. This can reduce confident fabrication, but it cannot guarantee the model will never make mistakes.

Practical skills: It was trained to write code, run programs, read and write files, and operate websites. The team uses complete tasks to demonstrate whether it can keep working through to completion.

Build a Job-Application Page, Then Fill In the Form

Inkling first generates a job-application webpage from a single request, including the role description, contact details, and reasons for applying. An embedded automation then reads the user’s saved profile and completes the form. The demo combines coding, interface generation, and browser operation.

Revise a Multiplayer Game Through 40 Rounds of Review

Another task was to build a multiplayer Snake game. The requirements covered server-authoritative game state, human and bot players together, collision rules, a leaderboard, tests, and startup instructions. Inkling built an initial version, then revised it through 40 rounds of GPT Codex review. The example demonstrates long-horizon task performance: the model can preserve context, take criticism, rerun tests, and keep correcting its work.

The fine-tuning demo has a specific goal: make all future English answers avoid the letter e. The model first writes a scoring function that gives 0 points if e appears and 10 points if it does not. It then generates training data and configuration, and uses Tinker to complete 32 batches over 3 epochs, for 96 training steps in total.

Technical Choices and Surprising Findings

Learning to Say Less: Internal Reasoning Turns Telegraphic

Inkling’s pretraining data contains text, images, audio, and video, totaling 45 trillion tokens. Post-training included more than 30 million reinforcement-learning task attempts. One attempt means the model completes a task, receives a score, and improves based on the result.

The training team found that the model’s internal reasoning automatically grew shorter as reinforcement learning continued. Early on, sentences were complete and used many connecting words. Later, the model dropped articles, prepositions, and some grammar, producing telegraphic fragments while reaching the same final answer. This behavior was not written into the reward objective; the model learned it under pressure to reduce token cost.

Replacing Mainstream RoPE with Relative Position Encoding

Many large models use RoPE, or rotary position encoding, to help remember where content appears in a long document. Inkling instead uses relative position encoding, focusing more on the distance between two pieces of content. Its attention layers also alternate at a 5∶1 ratio: five consecutive layers look at nearby information, then the next considers the full context.

How Can You Use It?

To try it now: Open the Inkling Playground on Tinker, currently free for a limited time.

To fine-tune or build on it: The full weights are on Hugging Face, along with an NVFP4 quantized version for NVIDIA Blackwell GPUs. To continue training with your own data, use Tinker and the official Cookbook.

Downloadable Does Not Mean It Runs on an Ordinary Computer

The full weights on Hugging Face occupy about 1.9TB. The BF16 version requires at least about 2TB of aggregate GPU memory, while the NVFP4 quantized version still needs at least about 600GB. Most developers are better served by the Playground, a third-party API, or Tinker. Self-hosting is a data-center-scale job.

The model card also lists a safety gap: harmful requests framed through role-play or indirect wording can occasionally bypass refusal behavior. Public deployments still need content filtering, rate limits, and monitoring.