Reflection, a startup backed by Nvidia, is preparing to release its first open-weight AI model, Axios reported on 4 October. The company is entering a market in which developers can obtain and adapt models themselves, while businesses have more often bought access to AI through services run by other companies.
Key points
- Reflection is preparing its first open-weight model, with no specific launch date or performance figures disclosed.
- Nvidia has invested $800 million in the company, according to WallStreetCN.
- Reflection wants businesses and governments to combine its models with their own data and computing resources.
- Other Western companies are expected to release open-weight models later in October.
Reflection’s model enters an uneven market
An open-weight model makes its trained numerical parameters available for download. Those parameters, or weights, are the settings a model has acquired during training: much as the positions of many controls determine what a machine does with an input, they determine how the model responds to a prompt. Developers can run and modify open-weight models within their licence terms, rather than relying solely on access to a service operated by the model maker.
Keeping sensitive work files under local control could be part of using such a system. In the arrangement Reflection proposes, a model would work with privately held data and computing resources on site, rather than only through an enterprise AI account.
The distinction matters because open weights have attracted individual developers and smaller businesses, while enterprises more commonly obtain AI through APIs or enterprise accounts. An API lets one system request an answer from a model supplied as a service. Running a downloadable model instead gives its operator a different way to combine the model with its own data, subject to the model’s licence and the computing it requires.
WallStreetCN described the most powerful open-weight models as almost exclusively Chinese, even though open models can account for much of the use on platforms offering a range of models. That is a different measure from enterprise adoption: the same report said open-weight models remain relatively limited in that market. Reflection’s proposed release would therefore enter a field with established models but room to seek business customers.
Nvidia’s $800 million backs Reflection’s plans
Nvidia has invested $800 million in Reflection. The chipmaker’s backing connects the planned model to the computing hardware Reflection wants to make part of its broader offering. Reflection calls that plan an “AI Factory”: businesses and government agencies would combine Reflection’s models with their own data and computing resources to build customised, lower-cost AI systems on their own premises.
That is a plan for deploying models, not a reported measurement of the model Reflection is preparing to release. WallStreetCN said Reflection had not disclosed a specific launch date or performance metrics. Reports cited by the publication expect the model initially to trail the most advanced US systems while potentially competing with leading Chinese open-weight models. Without published performance figures, that prospective comparison remains an expectation.
Reflection’s chief executive, Misha Laskin, previously worked at Google. In an interview with CNBC, he said reaching the leading edge would take time and compared building AI models with building large rockets. His caution sits alongside the company’s immediate ambition: to offer a model that organisations could use with data they already hold, rather than wait for it to match the strongest US systems.
Shinsegae’s 250-megawatt facility plan
In March 2026, Reflection signed a memorandum of understanding with South Korea’s Shinsegae Group for a 250-megawatt AI facility in South Korea. The planned facility would use Nvidia chips and run Reflection models. A memorandum of understanding records the parties’ stated intention to pursue the project; the facility is an example of how Reflection proposes to put its models and computing capacity together.
The proposed facility also gives Reflection’s on-premises idea a physical scale. Models made available for download still need machines to run them, particularly where an organisation wants to work with its own data. Reflection’s approach would pair access to the model with the computing required to operate it, extending the offer beyond a downloadable set of weights.
Other Western companies are expected to release open-weight models later in October. Reflection is preparing its first model amid those anticipated releases, while its computing plans extend beyond the proposed South Korean facility. In June 2026, the company signed a computing agreement with SpaceX that is due to run through 2029.
