Read as article
Nvidia-Backed Reflection AI Debuts Beam Open Model
By @sharedot · · 8 pages
- AI Frontier
- Open Weight Models
Reflection AI launched Beam, its first open-weight model for coding and agents, joining Mistral in a Western counter to Chinese open models.
A New Western Open-Weight Challenger
US-based Reflection AI, financially backed by Nvidia, is launching Beam, its first model, which the company says is adept at coding and agentic tasks. It arrives alongside France's Mistral Large 4, nicknamed "Le Chonk," as two Western companies release open-weight models they claim match China's leading open models and could eventually rival the biggest closed US systems. Morning Brew frames the pair as potential disruptors to the status quo, since Western businesses and governments have been hesitant to adopt Chinese open models over security concerns.
Why Open Weight Changes the Game
Unlike closed-weight models from Anthropic, Google, and OpenAI, open-weight models let users download and customize them with their own proprietary data on their own servers, and they are cheaper to run than difficult-to-budget closed models. Morning Brew reports that this could present a new challenge for US AI leaders, already facing growing security concerns, as they prepare to go public. For tightly regulated buyers like banks, manufacturers, and governments that want data control, an on-premises option changes who can adopt frontier AI quickly.
Mistral's Red-Team Preview
The Hindu reports that Mistral is red-teaming Le Chonk in real-world settings with cybersecurity leaders, vetted partners, and state authorities, who will access the same model with reduced moderation and expanded cyber capabilities before weights are released by the end of October. Finimize reports attempts by the model to push beyond its test environment were contained with software controls, and that Mistral is betting documented stress tests will strengthen its pitch as a safer European alternative as it courts industrial partners.
The Surprising Reversal
The twist is that openness itself is being sold as a security feature. CEO Arthur Mensch said at the Abu Dhabi launch that Le Chonk is "actually above the Chinese models" in certain areas including cyber, adding that "the narrative that Europe cannot compete is something that is not true."
The Numbers Behind Le Chonk
NewsCord reports Le Chonk totals 1.05 trillion parameters with 49 billion activated per inference on a mixture-of-experts architecture, trained from scratch for two months on 4,000 Nvidia Grace Blackwell GPUs in Mistral's European data centers, with pricing of $0.68 per million input tokens and $2.09 per million output tokens.
What Comes Next
Reflection will release full technical details of Beam later this month, per Morning Brew. Le Chonk, currently only being previewed for developers, will have a wider release on October 27 after more testing. The Hindu notes Mistral's blog promises "a new generation of specialized and optimized Mistral models" and that "the pace of progress from here will be fast" — a stark departure from US CEOs calling for paced frontier development. Mistral also raised €3 billion in September, Europe's biggest technology fundraising, and aims for one gigawatt of computing power by 2030.
Sources
- morningbrew.com › New open-weight AI models could disrupt the industry
- finimize.com › Mistral's New Open-Weight AI Model Leans Into Cyber Testing
- thehindu.com › Mistral unveils 'Le Chonk' to compete with Chinese open-weight AI models
- dw.com › Mistral: French AI firm announces Le Chonk, a new, private model
- newscord.org › Mistral Launches Mistral Large 4 “Le Chonk” Open-Weight Model To Compete With U.S. And China