As Chinese AI companies shake the global AI market with open-weight AI models that can be downloaded for free and even modified, startups in the United States are also emerging that tout open weights as their specialty.

A recent Wall Street Journal report said Arcee, Reflection AI, Poolside and Thinking Machine Lab are U.S. AI companies focused on open weights that are drawing interest. Nvidia is also expanding the ecosystem by releasing its own open-source models.

The global AI market is seeing Chinese AI companies such as DeepSeek, Moonshot AI, Z.ai and Alibaba expand share by offering open-weight models that deliver performance comparable to high-end models from OpenAI or Anthropic at very low prices. More U.S. companies are also using Chinese AI models.

The weights disclosed in open-weight models can be seen as data files that contain an AI model's "intelligence" and decision criteria. In large language models, weights play a key role in determining how AI responds to user questions or commands. For users, disclosed weight files can be downloaded, run directly on a PC or server and modified as needed. If weights are not disclosed, users have no choice but to use the model as designed by the AI developer.

ㆍ[Tech Inside] Everything about open-weight models shaking up the AI market

The current structure is one in which Chinese companies dominate, but the early open-weight model race was led by the United States. Meta led the early race with its Llama series but shifted its focus to closed models, and the open-weight model market was later reshaped around Chinese companies.

Against this backdrop, calls in the United States to grow open-weight models have recently been getting louder, and there is also movement to check Chinese open-weight AI models for geopolitical reasons. Some users want U.S.-made open-weight models because they worry about possible censorship or security risks in Chinese AI models, the WSJ reported. Chinese AI companies deny claims that they pose security risks, but concerns persist.

U.S. open-weight AI companies appear to be trying to expand their base on the back of this situation.

Jason Warner (제이슨 워너), co-founder and co-CEO of AI startup Poolside, which unveiled its open-weight AI model Laguna S 2.1 in July, told the WSJ that market demand is very strong for U.S. companies developing the best open-source AI.

Poolside said Laguna S 2.1 has 118 billion parameters. It showed performance comparable to or better than much larger models such as DeepSeek-V4-Flash, Nvidia Nemotron 3 Ultra and Thinking Machines' Inkling on the benchmark Terminal-Bench 2.1 and SWE-Bench Pro. It is not yet at the level of top-tier models, but already has sufficient performance, Poolside said.

Arcee, founded in 2023, started with a small model with 4.5 billion parameters and expanded to its 4.5 billion-parameter "Trinity Large" model.

The company said Trinity Large was trained on 2,048 Nvidia Blackwell B300 chips under a plan to make the most of a $20 million budget. Arcee previously raised $50 million in funding at a valuation of $240 million.

Mark McQuade (마크 맥쿼드), Arcee's CEO, said: "A more important goal was to prove that it could be done with only limited funding and resources."

Interest in open-weight models is growing in the United States, but more verification appears to be needed in terms of potential. Open-weight-based AI companies, in particular, are finding it difficult to raise funding compared with closed-model companies. Voices are growing that the United States must foster domestic open-weight AI to compete with China, but investors remain cautious about backing these startups, the WSJ reported.

McQuade said on social media platform X (Twitter): "If it is a venture capital firm talking about the need for U.S. open-weight AI, I have probably met them and pitched them." He said: "Virtually all of the top VCs said they would not invest. Now it has made me even more determined. We are not just talking, we are actually training and releasing models. Our goal is to help the United States catch up with China and ultimately get ahead."

It is also still uncertain whether they can become an alternative to Chinese companies in terms of performance. Reflection AI secured billions of dollars from investors including JPMorgan Chase, Sequoia Capital and 1789 Capital, where Donald Trump Jr. is a partner, including an $800 million initial investment from Nvidia.

In a recent funding round, it was valued at $25 billion. The Defense Department also included Reflection AI in an AI supplier group along with 7 domestic big tech companies. Even so, Reflection AI has not yet released its first model.

Reflection AI pivoted from developing coding tools to developing open-source AI models after meeting Nvidia CEO Jensen Huang last August, but it has not released a model for almost a year, The Information reported.

Reflection AI plans to release its first model within this year and then sequentially unveil a large model and a small model in early 2027. Citing sources familiar with the company, The Information reported the first model will match top-tier Western models at launch but lag leading Chinese open-source models.

Thinking Machine Lab, led by former OpenAI CTO Mira Murati (미라 무라티), also released its first open-weight model, Inkling, in July, scoring 41 points on AI evaluation firm Artificial Analysis' intelligence index. Kimi K3 scored 57 points, and Anthropic's latest model scored 61.