hina and Europe have typically led the charge on open models. Now, another US company is joining the fray with a significant entrant into the race.
On Wednesday, Thinking Labs, the AI startup founded by ex-OpenAI CTO Mira Murati, launched its first model: Inkling. The model sets itself apart from leading AI labs in the US in a very important way: it is open-weight, allowing customers to customize it directly to their needs through fine-tuning and training.
To suit that purpose, the model was built to perform across a wide range of areas. Or, as Thinking Machines describes it, it's a "generalist model" that can reason about text, images, and audio rather than being proficient in just one domain. This is evident in benchmark performance, which isn't best-in-class but is consistent across specialties, as shown in the image below.
Other model features include:
Parameters: Mixture-of-Experts transformer model with 975 billion total parameters, 41 billion active.** Context window**: Up to one million tokens** Pretrained**: On 45 trillion tokens of text, images, audio and video, allowing it to be proficient across all three.** Lightweight counterpart**: Inkling-Small, a lighter-weight model with 12 billion active parameters.** Fine-tuning**: To make customization accessible, the company says it is making Inkling available on Tinker, the company's training API platform.Cost-efficiency: Inkling's "controllable thinking effort" allows users to customize the cost/performance curve.
Until now, most open-source models have been developed in other countries, with China leading the way. These models are good for the ecosystem because they allow people and organizations to more deeply specialize the models, adjust their behavior, run them locally, and even save money. Though, as Thomas Randall, Research Director at Info-Tech Research Group, told The Deep View, they may not be for everyone.
"If organizations have use cases where cost control, sovereignty, fine-tuning, deployment flexibility, or domain-specific performance matter, then open-source models are relevant," said Randall. "For the highest-risk or most complex work (e.g., legal, medical, cyber, or regulated decisions), closed frontier model providers are preferable because of safety tooling, support models, auditability, and more mature deployment ecosystems."
Still, for the US to compete in the open-weights model space, it has a lot left to do, despite recent entrants from not only Thinking Machines but Nvidia, Google and OpenAI. A closer look at benchmark performance shows that many of China's leading labs, which have also taken a generalist approach with their open-weight models, have outperformed Inkling on benchmarks. That includes GLM 5.2, DeepSeek V4 Pro, and Kimi K2.6. Kimi K3 also just released on Thursday, and it's Kimi’s most capable model to date, with 2.8 trillion parameters.
Our Deeper View #
Competition is the lifeblood of any industry. This is especially true for the state of AI we are in now, where enterprises and AI builders are prioritizing cost, and open-weight models provide a tangible solution while applying pressure on the leading labs to be more cost-conscious. Just this month, OpenAI and Meta have both released new models, each highlighting their cost efficiency. Still, it is important to keep in mind the broader backdrop, including the rising costs of finite resources such as data centers and energy. Open-weight models still need the same compute power, even if they cost less than proprietary models.