DeepSeek V4.1 Flash Goes Live — Faster, Cheaper, Multimodal
DeepSeek is officially releasing its V4.1 Flash model today, September 10, capping a beta testing period that began just two days ago. The Chinese AI lab claims the new model comprehensively surpasses V4 Pro across performance, cost, speed, and total time metrics — a bold claim given that V4 Pro has been one of the most competitive models available.
The numbers are striking: V4.1 Flash processes up to 427 tokens per second, features a completely new model architecture with native multimodal support, and comes with significant price reductions. In a notable move, DeepSeek announced that until V4.1 Pro launches, all V4 Pro API requests will be automatically routed to V4.1 Flash — and billed at Flash's lower price point.
The release arrives at a particularly sensitive moment, as V4.1 Flash drops just days after U.S. intelligence agencies singled out DeepSeek in a major distillation advisory. Whether this timing is coincidental or strategic, the model's performance claims will face intense scrutiny from both competitors and regulators.
NSA, CISA, and FBI Accuse Six Chinese AI Firms of Industrial-Scale Distillation
Three of America's most powerful intelligence and cybersecurity agencies dropped a joint cybersecurity advisory on September 8 that names six Chinese AI companies — DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI — and accuses them of running “aggressive, malicious, and targeted” distillation campaigns against U.S. frontier models since late 2024.
According to the advisory (AA26-251A), these companies extracted billions of tokens across millions of queries from Anthropic's Claude, OpenAI's GPT, Google's Gemini, and xAI's Grok. The technique — knowledge distillation — involves using a more capable model's outputs to train a cheaper, smaller model, effectively transferring capabilities without the original R&D investment.
Perhaps most striking is the recommended mitigation: rather than blocking suspicious accounts outright, the agencies advise American AI providers to quietly degrade responses for accounts identified with high confidence as conducting malicious distillation. This “soft defense” approach is designed to make detection harder for the distillers while preserving normal service for legitimate users.
The advisory is likely to reignite the already heated debate over whether open API access and model distillation represent a fundamental national security risk or simply the natural flow of knowledge in a global industry.
Source: CISA, Northeast Times, Help Net Security
OpenAI's Navier-Stokes Claim: Historic Breakthrough or Attribution Scandal?
OpenAI announced on September 8 that an internal AI system — reportedly more capable than GPT-6 Astra — produced a proof of finite-time blowup for the forced 3D Navier-Stokes equations, one of the seven Clay Millennium Prize Problems, each carrying a $1 million reward. The company says roughly 10,000 coordinating AI agents exchanged 2.7 million messages and generated approximately 130 billion tokens over an 88-hour autonomous run, with Astra subsequently spending 17 hours formalizing and verifying the proof.
If validated, it would mark AI's most significant contribution to pure mathematics to date. But the announcement was immediately engulfed by an attribution controversy.
NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge had been collaborating on closely related research using AI tools. Buckmaster alleges that OpenAI's Sébastien Bubeck offered him sole authorship — on the condition that Alpöge's name be removed from the paper due to his Anthropic affiliation. Buckmaster says he refused, calling the offer an attempt to sideline his collaborator.
The controversy strikes at a trust question that will only grow as AI labs provide ever more powerful research tools: can researchers safely use frontier lab platforms to work on unpublished discoveries without those labs leveraging the data?
Source: Fortune, Axios, VentureBeat
Mistral AI Closes €3 Billion Round — Europe's Largest Ever
Mistral AI has raised €3 billion in a Series D led by Samsung Electronics, with co-leads including the Scaleup Europe Fund (managed by EQT) and PSG Equity. The round values the Paris-based company at over €21 billion post-money — nearly double its €11.7 billion valuation from just a year ago.
The investor list reads like a who's who of global tech and finance: Advent, BlackRock, the Grand Duchy of Luxembourg, a16z, ASML, Bpifrance, NVIDIA, and Korelya Capital. It is the largest equity funding round ever completed by a European technology company.
The Samsung partnership carries strategic weight beyond the capital. Samsung will integrate Mistral's technologies, including its flagship Mistral Large model, across its semiconductor operations. For Mistral, this represents a major hardware partnership that could give it a compute and deployment advantage in the increasingly capital-intensive AI race.
Mistral plans to use the funds to scale compute capacity, build infrastructure, accelerate commercial growth, and expand internationally — framing its mission around “sovereign, open-weight AI” as an alternative to the U.S.- and China-dominated model landscape.
Source: CNBC, TechCrunch, Mistral AI
Model Fatigue: Too Many AI Models, Too Fast?
After Anthropic, OpenAI, Meta, and Google all shipped major new models in the first week of September — Fable 5.1, GPT-6 Astra, Muse Spark 1.3, and Gemini 3.8 Flash respectively — CNBC reports that a new phenomenon is setting in among enterprise buyers: model fatigue.
Runpod CEO Zhen Lu described the pace as “disorienting” for IT buyers trying to make procurement decisions. Meanwhile, more than 1,100 employees across major AI labs have signed a petition asking Washington to help pace frontier AI development — a remarkable move from within the industry itself.
The fatigue is compounded by McKinsey research showing that nearly one-third of surveyed organizations have decided against purchasing software products because they could build equivalent functionality using AI coding agents. The line between “AI models” and “AI-built software” is blurring, making purchasing decisions even more complex.
Source: CNBC