MiniMax's 2.7T model, SambaNova's $1B raise, and a nasty new LLM exploit

MiniMax plans a 2.7T open-source model, SambaNova hits $11B valuation, and 'HalluSquatting' turns LLM hallucinations into botnet fuel.

The big picture

The hardware and model layers of the AI stack are both heating up fast — bigger open-source models, fresh chip funding, and a new inference speed play — while a clever new attack vector reminds us that LLM quirks aren’t just annoying, they’re exploitable.

The open-source model arms race gets a new entrant

Chinese startup MiniMax has announced plans to open-source a 2.7 trillion parameter model later this year. For context: that would make it one of the largest openly released models ever, dwarfing Llama 3’s 405B and putting it in territory that only a handful of closed frontier labs have reached at all. (The Decoder)

The announcement is worth watching but comes with an obvious caveat: “plans to release later this year” is doing a lot of work. MiniMax hasn’t shipped this model yet, and parameter count alone tells you nothing about actual capability or efficiency. Still, if they follow through, a 2.7T open-source model changes the calculus for anyone building on open weights — suddenly, frontier-scale becomes self-hostable. That’s a big deal for enterprises that can’t send data to an API.

AI inference and chip infrastructure: money is still pouring in

SambaNova, which makes purpose-built AI inference chips, closed a $1 billion Series F at an $11 billion valuation — just five months after its previous mega-round. This comes after earlier reports that Intel had tried to acquire the company for roughly $1.6 billion, which in retrospect looks like a lowball offer SambaNova was right to decline. (TechCrunch)

The pace of this fundraising — two enormous rounds inside half a year — signals that inference-optimized hardware is where serious capital is going right now. Training compute is mostly spoken for by the hyperscalers; the open question is who wins the inference layer as model usage scales. SambaNova is betting that commodity GPUs won’t cut it.

On the software side of inference, French startup ZML released ZML/LLMD for free. The pitch: a single runtime that accelerates inference across a wide variety of AI chips, not just Nvidia’s stack. Yann LeCun (Turing Award winner, Meta’s chief AI scientist) has publicly backed the company, which gives it credibility beyond typical startup noise. (TechCrunch)

The hardware-agnostic angle is genuinely interesting — anyone running on AMD, Intel, or custom silicon currently has a rough time getting competitive inference performance. If ZML/LLMD actually delivers, it could matter a lot for teams that don’t want to be locked into Nvidia’s ecosystem. The “free” release is also a smart go-to-market move: get developers hooked before monetizing.

HalluSquatting: when LLM hallucinations become a weapon

Security researchers have found that nine major AI tools can be exploited to build botnets using a technique called “HalluSquatting.” The attack works by weaponizing LLMs’ tendency to hallucinate package names or URLs that don’t exist — attackers register those hallucinated domains or package names in advance, then wait for developers (or AI agents) to blindly fetch from them. (Ars Technica)

This is one of the more creative attack vectors to emerge from the LLM era, and it’s particularly insidious because it doesn’t require breaking anything — it just exploits a model behavior that everyone already knows exists. If you’re using AI-generated code suggestions or agent workflows that make network calls, you need to validate every package name and URL before executing. Treating AI output as trusted input here is the mistake. Expect this class of attack to grow as agentic systems proliferate.

Meta’s always-on glasses prototype raises obvious questions

Meta is internally testing AI glasses with a “Super Sensing” mode that uses an always-on camera and microphone to log everything the wearer sees and hears throughout the day. This is a prototype, not a product announcement. (The Decoder)

The capability is technically impressive and socially complicated in ways Meta seems to be exploring with some care — or at least testing internally before inflicting on the public. For developers building on Ray-Ban Meta’s platform, this is a preview of where the API surface could expand: continuous context, persistent memory, ambient awareness. Whether any of that becomes a real product depends entirely on whether regulators and users tolerate it.

Quick hits

Former OpenAI CPO Kevin Weil has joined the board of Stoke Space, a reusable rocket startup — ex-AI-exec talent is now flowing toward aerospace, apparently. TechCrunch
Ars Technica ran a piece on general-purpose robot autonomy featuring robotics researchers and founders explaining where AI-driven physical automation actually stands today — worth a read if you’re curious how close “robot coworker” really is. Ars Technica