Nvidia and Reflection Push Open AI Models Into Enterprise

The open AI models race flipped in one crowded week. Nvidia and Reflection AI released powerful new open AI models built for enterprise work, answering months of momentum from Chinese labs — and DeepSeek and MiniMax fired back with price cuts of their own. According to reporting from PYMNTS, the market for open AI models is no longer a Chinese monopoly, which matters for banks and insurers that could not run Chinese weights for security reasons.

For enterprises, the shift is practical: open AI models can be downloaded and run on a company’s own servers, keeping sensitive data in-house instead of sending it to a cloud API. That single property is turning open weights from a hobbyist option into a procurement conversation.

The American Open Stack Goes Enterprise

Nvidia released Nemotron 3.5 Super VL on October 15, an open 120-billion-parameter model that reads documents, charts and web pages, according to the company’s announcement. Nvidia said it runs five times faster than comparable open AI models. Spend-management fintech Ramp is already tuning it for financial document analysis, and Scale AI trained a version for a client’s merger due diligence work. Nvidia’s pitch is simple: a bank could review credit files against its own lending rules without data ever leaving its servers.

Two days later, Nvidia-backed startup Reflection AI released the full weights for Beam, a 501-billion-parameter model, under a free commercial license. Reflection first previewed Beam two weeks earlier as a U.S. answer to Chinese open AI models, and it launched with more than 40 partners, including AWS, Google Cloud, Microsoft and Oracle, according to Reflection AI. Dell and HPE will ship it for on-premises and air-gapped environments — the kind of deployment regulated industries demand.

The Middle East entered the race too: the Institute of Foundation Models and Inception42 released Jais 3 on October 15, a family of six open Arabic models ranging from 900 million to 375 billion parameters, free to download, plus Jais Speech for Arabic speech recognition.

China Answers With Price Cuts

DeepSeek updated its low-cost Flash model on October 17. The Chinese lab said the new version scores higher on every benchmark it reported, finishes tasks faster, and writes shorter answers — which lowers the cost per job. The weights are free to download under an MIT license, the most permissive kind. DeepSeek has been a defining force in the open AI models market all year, as covered earlier after its massive funding round.

Shanghai-based MiniMax launched M3.1 Flash on October 19, saying the model runs nearly three times faster than its predecessor while cutting usage costs by 70 percent. It scored 76.2 on BankerToolBench, a test of multi-step financial tasks and compliance work, at pricing of $0.15 per million input tokens and $0.60 per million output tokens. The weights will be released later. For agent builders, the economics matter more than the leaderboard: lower inference costs make AI agents viable for more applications, and the real metric is the total cost of completing a task reliably.

Speed Becomes the New Benchmark

Inception Labs introduced Mercury 3 on October 19. The startup builds models on diffusion — generating text in parallel instead of one word at a time — and said Mercury 3 matches or beats OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5 while running four to eight times faster. Support-agent builder Lorikeet told Inception the model beat rivals on both speed and accuracy in early testing.

Perplexity released two open decision models the same week. Decision models don’t write text; they pick from a fixed set of options and return a probability. The new versions answer several questions about the same input in one pass, which suits tasks like routing tickets or screening transactions — exactly the kind of fast, cheap decisions that agent workflows make thousands of times a day.

The shape of the market is changing. Open AI models now arrive in three flavors — U.S. enterprise-grade, Chinese low-cost, and speed-optimized — and each has a real distribution channel, from cloud marketplaces to air-gapped servers. The shift from API to self-hosted also changes the competitive dynamic of cloud AI: enterprises can now blend managed model access, like the OpenAI models now available on Google Cloud, with open weights they run themselves. For banks, insurers and governments, the choice of what to run in-house just got a lot wider.