Mira Murati's Thinking Machines Lab (@thinkymachines) released Inkling on Wednesday, July 15th, giving the heavily funded AI lab its first trained-from-scratch foundation model and making the full weights available for developers to fine-tune and deploy, according to a July 15th announcement and a six-post thread on X.
Inkling is the first concrete model release from Murati's post-OpenAI startup after a year in which Thinking Machines drew attention mostly for the scale of its team and financing. WIRED reported in July 2025 that Thinking Machines raised a $2 billion seed round at a $12 billion valuation led by Andreessen Horowitz, with Nvidia, Accel, Cisco and AMD among the investors. The bet behind Inkling is narrower and more operational than the size of that round suggested: Thinking Machines is trying to win developers that want an adaptable base model, rather than another closed general assistant rented through a frontier API.
Murati, who was OpenAI's CTO before leaving in September 2024, has framed Thinking Machines around accessibility, customization and multimodal collaboration. That shows up directly in Inkling. Thinking Machines says Inkling reasons across text, images and audio, supports controllable thinking effort for tuning latency and cost, and is available for fine-tuning on Tinker, the training API Thinking Machines launched in October 2025.
Inkling is open-weight, but the pitch is control
Thinking Machines is not presenting Inkling as the strongest model in the market. In its own release, Thinking Machines says Inkling "is not the most performant model available today, closed or open." That caveat is the point. Thinking Machines is selling Inkling as a practical model for teams that need to adapt behavior, manage cost, and run specialized workflows where a closed API may be too expensive or too rigid.
The model is a mixture-of-experts transformer with 975 billion total parameters and 41 billion active parameters, according to Thinking Machines. It supports a context window of up to 1 million tokens and was pretrained on 45 trillion tokens spanning text, images, audio and video. Thinking Machines says Inkling uses 256 routed experts and two shared experts per MoE layer, with six routed experts active per token. The architecture is built around efficiency and long-context performance, with a mix of sliding-window and global attention layers.
The weights are available on Hugging Face under an Apache 2.0 license, including the original checkpoint and an NVFP4 checkpoint for Nvidia Blackwell systems. The Hugging Face model card describes Inkling as a general-purpose multimodal model that accepts text, image and audio inputs and generates text outputs. It is intended for developers building agentic systems, coding assistants, chatbots, retrieval-augmented generation products and other applications.
That license and distribution strategy put Inkling into the open-weight lane at a time when enterprises are weighing the cost and control tradeoffs of closed models. The weights are open; the commercial path still runs through Thinking Machines. Inkling is available on Tinker with 64K and 256K context options, and Thinking Machines says full pricing is posted in its Tinker documentation. The company is offering a 50% discount for a limited time, according to the launch post.
Tinker turns Inkling into a customization product
Inkling matters most as an extension of Tinker. Thinking Machines describes Tinker as a training API that lets researchers control model training and fine-tuning while Thinking Machines handles distributed infrastructure. The platform exposes primitives including forward_backward
, optim_step
, sample
and save_state
, and uses LoRA so multiple fine-tuning jobs can share the same compute pool.
Thinking Machines used the Inkling launch to show a self-fine-tuning demo: Inkling wrote a fine-tuning job, ran it on Tinker, evaluated the result, and switched to the updated weights. The example task was intentionally narrow, forcing the model to respond without using the letter "e." The demo finished after roughly 27 minutes, according to the company. That is a toy example, but it explains the product direction: Thinking Machines wants developers to treat a foundation model as something they can reshape in the workflow, rather than a fixed remote service.
The Tinker positioning also gives Murati's lab a cleaner answer to the business question around open weights. Releasing weights can dilute API lock-in, but it can also create demand for hosted fine-tuning, deployment support and managed training runs. Thinking Machines is pairing the model release with hosted customization, a playground in the Tinker console, cookbook updates, audio fine-tuning recipes in the Tinker Cookbook, and a tml-renderers package for sampling and post-training with chat templates, tool calls, reasoning content and multimodal inputs.
Thinking Machines also lined up day-one distribution. Inkling is available through APIs from Together AI, Fireworks AI, Modal, Databricks and Baseten, according to the launch post. Thinking Machines says it worked with RadixArk on support for SGLang and Miles, Inferact on vLLM, Lightseek on TokenSpeed, Unsloth on llama.cpp, and Hugging Face on transformers integration.
Audio is the technical emphasis
Inkling is broadly multimodal, but Thinking Machines is putting particular emphasis on audio. The lab says Inkling transcribes speech, follows spoken instructions, answers questions about recordings and reasons over longer audio. On company-reported benchmarks, Inkling scores 91.4% on VoiceBench, 77.2% on MMAU and 56.6% on AudioMC. Those scores trail Gemini 3.1 Pro in the table Thinking Machines published, but Thinking Machines claims Inkling ranks among the strongest open-weight audio models.
The model's audio and vision path is encoder-free, according to the company. Audio signals are input as discrete dMel spectrograms, while images are encoded as 40-by-40 pixel patches using a four-layer hMLP. Thinking Machines says both are projected into the shared token stream and processed jointly with text. During inference, Inkling can use Python for visual tasks such as zooming and cropping, which brings image analysis closer to tool-using agent behavior.
Thinking Machines also published a safety and limitations section with the Hugging Face release. The model card says Inkling may hallucinate, fail to follow instructions precisely and degrade in long multi-turn conversations. It also says Thinking Machines found Inkling did not present material capability uplift beyond what is already available in the open-weight market, while warning downstream developers to use application-level safeguards and human oversight for high-stakes uses.
The first in a family
Thinking Machines is also previewing Inkling-Small, a lighter MoE model with 276 billion total parameters and 12 billion active parameters. Thinking Machines says Inkling-Small was trained with a similar recipe and, in its internal tables, matches or exceeds the larger model on some math, science, vision and audio benchmarks. The full weights are not available yet. Thinking Machines says it is finishing testing and will release them later.
That sequencing gives Murati's team room to test the market with a large open-weight release while preparing smaller variants for lower-cost workloads. Thinking Machines names coding, model grading and synthetic-data generation as use cases where a smaller model with controllable thinking effort may be a better fit.
The release also clarifies what Thinking Machines has been building since Murati assembled a team of former OpenAI researchers and other AI infrastructure veterans. The first model is not a consumer chatbot. It is a base model, a fine-tuning product and a distribution push packaged together. Inkling gives developers the artifact they can run, while Tinker gives Thinking Machines the service layer where customization, compute and customer relationships live.