You could download a trillion-parameter model on Thursday and run it on a single Mac, if that Mac has half a terabyte of memory. Hugging Face's own test clocked it at fifteen tokens a second. So, technically local. In practice, one coffee break per paragraph. This is AI Daily Briefing for Friday, October 30th. Coming up: Mistral puts the weights for Large 4 on Hugging Face and pitches it to security teams, Canada's government buys into Cohere's record round, Alphabet's stock drops after Google admits an AI Mode query earns less than a web search, and an Indian neocloud turns to the debt markets to pay for Nvidia's Rubin chips. Hit follow and you won't have to come looking for the next episode.
Aritra Roy Gosthipaty, writing on the Hugging Face blog about Mistral Large 4:
Mistral Large 4 scores 82% on the Artificial Analysis Cyber Index test that asks a model to reproduce a real vulnerability in open-source software and then patch the software. Mistral reports that several closed frontier models score near zero on the same test because they refuse to attempt the work.
We first talked about Large 4 on October 10th. What changed Thursday is that the weights landed on Hugging Face, in FP8 and NVFP4, with day-zero support in transformers, vLLM and llama.cpp. It's a mixture-of-experts model with about a trillion total parameters and a one-million-token context. Hugging Face says forty-nine billion active; the checkpoint name and other summaries say fifty-two. The cyber pitch is the part I care about. Reproduce, then patch, is how defenders actually work, and if your API provider refuses the reproduce step in the middle of an incident, you're stuck. Gosthipaty also notes its refusal rate on harmful cyber prompts is higher than any other open-weight model's. So this isn't an unfiltered model either. Keep the rankings in proportion, though. Hugging Face calls it the fifth-best open-weights model on the Artificial Analysis index. The French outlet ActuIA reports its score fell from thirty-eight to twenty-five when Artificial Analysis moved to version five, the rebuild we covered yesterday. And ZeroHour's summary of Latent Space says Mistral's ranking claims are narrower than the independent scores. Then read the license. It's the Mistral Ecosystem Open-Weight License, an MIT derivative with limits on model-as-a-service offerings. Self-host it, fine-tune it, go ahead. Resell it as an API, call a lawyer first. And the realistic deployment is Hugging Face's Inference Endpoints recipe: eight H200s, about sixty-five tokens a second on a single request. That's a serious box for the fifth-best open model. Builders noticed anyway. The Hacker News launch thread passed thirteen hundred points. Meanwhile Kimi K3.1 Pro's weights are still promised for November 4th.
Joe Castaldo and Murad Hemmadi, writing in The Globe and Mail:
To secure Ottawa’s investment, Cohere has committed to spend $1.7-billion on the buildout of compute infrastructure in Canada by 2030. It has also promised to remain headquartered in the country.
Cohere announced Thursday that it has secured most of a planned seven-billion-dollar round at a thirty-five point seven billion valuation, in The Globe's figures, the largest financing ever by a Canadian startup. Ottawa puts in up to one point six billion for an equity stake, paid out as Cohere hits hiring and infrastructure targets. Schwarz Group, the German retail giant, leads. Nvidia, Inovia and Saudi Aramco's Prosperity7 return. Watch the step-up. Last month The Globe reported Cohere was in talks to raise up to three billion US at a twenty-billion valuation. The round last September valued it around seven billion US. CFO Francois Chadwick says demand ran past the original plan, so the round stays open until it closes. And Ottawa is buying a commitment to stay put. Prime Minister Mark Carney said Thursday that under the agreement, Cohere will stay Canadian. The company also agreed to double a headcount that will top a thousand after its Aleph Alpha merger. Now the part the press conference skipped. The Globe says Cohere's models significantly lag OpenAI and Anthropic on coding and cybersecurity in industry rankings, and Aidan Gomez has acknowledged the gap. The last revenue number it gave investors was a two-hundred-forty-million-dollar annualized rate for 2025. Sovereignty is a real sales pitch to banks and governments. It doesn't close an eval gap. The investors aren't paying for leaderboard wins. They're paying for a supplier to regulated industries, finance, telecom, health care, energy, that is neither American nor Chinese. Whether that market supports this valuation is the bet.
Ars Technica, on what Google told investors Wednesday:
The other issue is that Google’s take on AI search isn’t pulling in revenue like standard web search yet. According to Schindler, each AI Mode query earns less revenue compared to a similar web search—there are fewer ads, and they cost less to run.
Yesterday we gave you Alphabet's top line, revenue up twenty-five percent. The market looked past it. Ars reports the stock was down almost twelve percent at publication, one of the largest post-earnings drops in the company's history. Earnings per share came in at two dollars twelve against a consensus of three oh two, and net income was almost nine billion lower than a year earlier. The search numbers are where it hurts. Search volume grew twelve percent, down from seventeen in Q2 and nineteen in Q1, and search revenue was flat. Then Philipp Schindler says the AI answer earns less than the blue links it replaces. And Google still plans to spend even more on AI next year. Ars's headline says the fix is more ads. If your product lives on Google traffic, ads inside AI answers is the change to watch. Cloud was the bright spot. Ars says it crossed thirty billion, about a billion over analyst estimates. Amazon closed out the week on Thursday. TechCrunch reports Andy Jassy said AWS will be a trillion-dollar business, but that it has to spend big first. ZeroHour's summary has Amazon still expecting about two hundred twenty billion dollars in cash capex this year, mostly for AWS, and more in 2027. So the sold-out-through-2027 line we aired Monday comes with a price tag. Google and Amazon both just told investors the same thing. Spend first, collect later.
Priyanka Salve, writing for CNBC on a Bloomberg report:
The split between bonds and loans is yet to be finalized, but the firm expects to get $300 million in advance payment from potential offtakers, the report said.
The firm is AM Intelligence, a Hyderabad company owned by the founders of the renewable energy firm Greenko. Bloomberg reported Friday that it wants up to three billion dollars in bonds and bank loans to buy Nvidia Rubin GPUs, and the company claims to be among the largest committed Rubin buyers outside North America. It didn't respond to CNBC. On October 5th it announced a six-billion-dollar plan for a hundred megawatts across India and Malaysia, twenty-nine thousand Rubin chips. Run the numbers. Six billion for a hundred megawatts. Then three hundred more megawatts in the next fifteen months for over twenty billion. Against a three-billion-dollar debt raise, the offtakers are putting up three hundred million in advance, and nobody has named them. It's the pattern we've followed all month: compute bought with borrowed money and backed by customer commitments, from Naver's Brookfield deal to CoreWeave bonds yielding over fifteen percent. Counterpoint's Neil Shah tells CNBC that Indian energy companies, including Adani, Reliance and Tata Power, are turning into neoclouds that lease capacity to hyperscalers. Power companies as GPU landlords makes sense. They already have the land and the electrons. What they don't have is a track record pricing GPU depreciation. I'd want the names on those offtake contracts before I touched the bonds.
Mistral's model card makes a safety claim this week: Large 4 refuses harmful cyber prompts more often than any other open-weight model, and open weights let researchers check that for themselves. Testing claims like that is the job of our sister show, AI Safety Daily. Every weekday it covers alignment, model evaluations, emerging risks and governance: what changed, what the evidence says, and why it matters. Look up AI Safety Daily in your podcast app.
My weekend takeaway: the trillion-parameter model is a free download, and the hard part is still the eight GPUs it takes to serve it. And the money keeps finding new ways in, from a national treasury to a bond desk. That's the week. AI Daily Briefing is a Lantern Podcast, and we're back Monday, November 2nd.