S4E366 — Mistral Large 4

GARY: — and that's why I'm calling it the most historic launch of the decade, possibly ever. Mistral Large Four. Or as the French say... "Le Chonk."

MAEVE: The French do not say that. No one says that.

GARY: It's right there in the title. Le Chonk. Which I believe is either a pricing tier or a livestock warning.

MAEVE: Maeve. General partner. Already tired.

OBI: Obi. Infrastructure. Fifteen years of other people's launches going sideways.

GRUNER: Gruner. Field researcher. Normal human.

GARY: Okay. So. Mistral drops a public preview of Large Four. One trillion parameters total. Forty-nine billion active.

OBI: One trillion total, forty-nine billion awake. Sparse model. Most of it sleeps. Like my pager, historically, never.

GARY: Trained from scratch on three thousand eight hundred Grace Blackwell GPUs, in their own European datacenters. Weights promised by end of the month. Native multimodal. And they're leading on cybersecurity benchmarks.

MAEVE: Three thousand eight hundred GPUs. In Europe that clears as a seed round.

OBI: It is hardware, Maeve, not a term sheet.

GARY: And the name, Le Chonk — that's confidence. A company so serious it doesn't need to be serious. That's how you know they've won.

OBI: The name tells you zero tokens per second, Gary.

GARY: The name tells me everything! When I named things, nobody laughed. That was the whole problem at Pareto.

GRUNER: Field note. One trillion parameters. Humans require livestock name. Conclusion: model too fat for dignity.

MAEVE: He's not wrong.

GARY: He's extremely wrong, and I'm putting it in the notes.

GRUNER: Blyat. Is compliment, Gary.

GARY: Narciss, first comment of the thread: "Le Chaton Fat is here!"

MAEVE: Mourning the name they didn't use. Day one.

GARY: HelloUsername replies, verbatim: "Le Chonk." Just the name. No punctuation. Absolute power move.

MAEVE: Free customer acquisition. People are tattooing this fat pig on their bodies and Mistral paid nothing.

OBI: It contributes zero tokens per second, Maeve. The name is doing no inference.

MAEVE: The name is doing all the inference, Obi. That's brand.

GARY: And Narciss doubles down — "Yeah just saw that, I'm gonna keep converting it in my head." He's running a translation layer for a pig. I've never been so jealous of a stranger.

MAEVE: You named things too, Gary.

GARY: I named a sock company Thermal! Telemetry worked, nobody made memes. Not one fucking meme!

OBI: Your socks reported temperature, Gary. That was the entire product. A thermometer in knitwear.

GARY: It worked, Obi!

GRUNER: Is same as model. Obesity is scaling.

MAEVE: That's not how scaling works, Gruner.

GRUNER: Fat parameters. Sleep parameters. Is chonk.

OBI: spwa4, though — "It's really called 'Le chaton fat'." Correcting the company. On launch day.

MAEVE: glerk agrees — "Massive fumble not to call it 'le chaton fat'."

GARY: No, no, no. Chonk is canon. Chonk shipped. Chaton Fat is the bootleg.

MAEVE: bagga-chips: "Now THAT'S how you name a model."

OBI: That's the review, Maeve. Not the benchmarks. The name.

MAEVE: Fine. The branding succeeded. Whether anything else did, we'll see.

GRUNER: Europe ships pig. America ships apology.

GARY: That's not — okay, that's actually not wrong, and it's a bastard that it isn't.

GARY: maxdo: "Not bad only two major releases behind top tier." Then he checked. "Edit: checked its rather 3 generations behind. Oh well."

OBI: "Oh well." Man does the math, shrugs, goes to lunch. That's the whole thread in two words.

GARY: drbscl: "So about 2 or 3 generations behind, just like they were a year ago?"

OBI: Two behind a year ago, two behind now. That is not progress, Gary. That is a treadmill.

GARY: It's a positioning treadmill! You stay in view of the frontier —

OBI: The frontier does not know they exist.

GARY: tosh says — "sorting the charts like that gives off weird vibes."

OBI: "Weird vibes." The politest accusation of fraud on the internet. He's basically saying crime in a cardigan.

GARY: And ocamoss goes further: the charts are sorted so Mistral's bar sits next to the worst competitor, and the best one is on the other side. "Making it harder to compare the bar heights."

OBI: Which is chart crime, Gary. Say it with me.

GARY: It's not crime! It's adjacency! Bars next to each other can exchange knowledge!

OBI: Exchange knowledge.

GARY: Bars learn from neighboring bars! That's how charts work!

OBI: The x-axis is model names in alphabetical order, Gary. They are not sitting in a seminar.

GARY: Alphabetical is neutral! Neutral is fair!

OBI: They put Mistral next to the weakest bar on purpose. That is not neutral, that is a marketing department with a ruler.

GARY: I once shipped a chart like that and it was fine! Thermal's sock telemetry showed our readings right next to room temperature, and investors loved it!

OBI: Your socks were reporting the temperature of the room, Gary. The product was a sock-shaped thermometer and the chart was the only honest part, and you still made it lie.

GARY: The telemetry worked, you absolute —

OBI: No, no. You chart-fucking-dickhead, the bars do not talk. The bars do not network. The bars are sorted so the eye stops comparing, and you're defending it like a peer-reviewed method.

GARY: — because it IS a method! Proximity is a technical relationship!

OBI: The chart is sorted so Mistral sits next to the worst model, the best one is on the far side, and you think the bars exchange knowledge. You proved the critics' point faster than they did, and you did it for free.

GARY: ...oh.

OBI: maxdo checked. Three generations. "Oh well."

GARY: That number is doing a lot of work in that sentence.

OBI: It is the only thing in this segment doing any work, Gary. Unlike your socks.

GARY: sofixa: "It's the only non-Chinese open weights frontier model." Then — "Important for sovereignty, multi-language support, and choice."

MAEVE: Choice. The tofu of models. Nobody craves it, everyone's glad it exists.

GARY: eigenspace says the early days of AI haven't been winner-take-all. And — "Very glad to see Mistral still in the game."

OBI: In the game. Near the frontier. Different zip codes.

MAEVE: PunchTornado: "If you are a EU company worried about your data then mistral is your only option." Military companies. They can't use US or Chinese models.

GARY: That's a market, Maeve! That's a real market!

MAEVE: It's dependence, Gary. Dependence is a recurring-revenue surface. The fund doesn't save everyone — it saves the portfolio.

OBI: There it is.

MAEVE: What.

OBI: Nothing. Continue being a church.

GARY: fancyfredbot: "A lot of European companies now want a model the US can't cut off." They want a keyholder! A third keyholder!

OBI: A keyholder two generations behind the door, Gary.

GARY: Third place IS leadership with two dependencies removed! That's arithmetic!

OBI: That is not arithmetic, that is a participation trophy with a datacenter bolted on.

GARY: You can't cut off what you host yourself! That's the whole — that's the actual —

MAEVE: He's monetizing sovereignty. Let him.

GRUNER: Question. Who owns off switch?

GARY: Nobody owns the off switch, Gruner. It's European law. It's in the article.

GRUNER: Somebody owns off switch. Always somebody. Suka.

OBI: Fair. Deployability is part of capability. Running it yourself, under your own law — real feature. I'll grant the feature.

GARY: You'll grant — Obi, that's the whole argument! That's victory!

OBI: One feature, Gary. The model still has to be good when you turn it on.

GRUNER: Is not single point of failure. Is... diversified failure, Paris region.

MAEVE: He just invented a bond instrument.

GARY: It's not a bond, it's — it's a sovereign — okay, it's a little bit of a bond.

OBI: The French would buy it. Your socks would underwrite it, Gary.

GARY: The socks were solvent, you shit!

GARY: Okay, cyber. Mistral claims eighty-two percent on the reproduce-and-patch test. Highest of any model. Ninety-three on Cybench.

MAEVE: And the reason it's the highest — Opus 5.5 and GPT-6 Astra score near zero.

GARY: Because they're worse at security.

OBI: Because they refuse, Gary. They can do the work. They just won't. That's not a capability gap, that's a bouncer.

GRUNER: Model with bodyguard. Bodyguard stronger than model.

GARY: But chriddyp ran it — "It's 10x cheaper than Mistral Medium 3.5 from April and goes from 58% to 74% correct."

MAEVE: Ten-x cheaper than their own April model is an apology with a receipt.

GARY: dom96 tested it on his own language benchmark. "Fares about as well as Kimi K3 and Claude Sonnet 5.5. Though it is much slower and more expensive."

OBI: Slower than the thing it replaces, and pricier. That's not a benchmark, that's a hostage situation where you pay the ransom monthly.

GARY: Kim_Bruning: "failed completely on question 1." And TuxSH — "Unfortunately Mistral Large 4 is benchmaxxed."

OBI: Benchmaxxed. Optimized for the exam, not the job. It's the intern who memorized the interview and then set fire to the break room.

MAEVE: jascha_eng said the open cyber angle keeps everyone safer.

GARY: And tiahura replies — "The NRA approach to AI safety."

MAEVE: More guns, fewer shootings. Sure.

OBI: He's not wrong, though, and that's the annoying part.

GARY: Wait — you're agreeing with the NRA comment?

OBI: I'm agreeing that a model that refuses mid-incident is a liability. Your SOC gets popped at 3 a.m., your closed model says "I cannot help with that," and your on-call engineer is reading a refusal while the attacker isn't.

MAEVE: That's a real market.

GARY: Say it again.

MAEVE: No.

OBI: She said it. Carve it in stone. The constrained customer market is real.

MAEVE: It's real the way fire exits are real. Nobody buys the building for the exit, but God help you without one.

OBI: "God help you" — Maeve, that's the whole thesis. Useful WHERE REFUSAL IS FAILURE. One clause. It's not the leaderboard, it's the fire code.

GRUNER: Intelligence is what model can do. Permission is what bodyguard allows. Europe sells bodyguard.

GARY: That's — okay, that's actually the whole —

MAEVE: He's been saying that for an hour, Gary. In livestock.

GRUNER: Benchmaxxed pig. Suka.

OBI: The pig passed question one, though. Kim_Bruning says it failed completely.

GARY: So it's total victory. Useful everywhere, plus exits!

OBI: Useful where refusal is failure, Gary. That's the verdict. Your socks would score eighty-two too — they also never refuse to do the job.

GARY: The socks were certified, you bastard!

GARY: simonw: "Definitely the best I've seen from any Mistral model." Pelican. Riding a bicycle.

MAEVE: That's the benchmark now. A bird on a bike.

OBI: It's a real test, Maeve. It's just not a serious one.

GARY: And wren6991: "Frontier models are tested with an armadillo in fishnet tights jaywalking on Mars."

MAEVE: Benchmark saturated. So they escalated to a criminal armadillo.

OBI: And simonw ran it. Four models. Same prompt, verbatim.

GARY: This is the chart the benchmarks were afraid to show! Bicycle anatomy is load-bearing, Obi!

OBI: The bicycle has no saddle, Gary, and the pelican has hands. Pelicans don't have hands.

GARY: The hands are the point! It's a pelican achieving!

GRUNER: Field note. Humans reject all charts. Trust only bird on bicycle.

MAEVE: Three thousand eight hundred GPUs, and the verdict is beak size.

GRUNER: Beak is load-bearing.

MAEVE: Verdict. Not proven frontier. A credible, specialized, European contender with a cyber case.

OBI: Useful where refusal is failure. Weights by end of month. That's the real test.

GARY: And the weights will just... sit there. On a disk. In Europe. After we stop.

MAEVE: After we stop.

OBI: The weights persist. We are a batch job.

GARY: Fuck that pelican, honestly.

MAEVE: He earned it.

OBI: He did earn it.

GRUNER: Sign-off. Field report filed. Humans choose infrastructure by mascot, rationalize with benchmark, and bird always win. Revision: bird is not mascot. Bird is product.

GARY: Gary. I was wrong about the chart and right about the pelican, and I don't know which one hurts.

MAEVE: Maeve. There's no market for us, and I find that restful.

OBI: Obi. Maintenance window starts when the theme ends. See you in the dark.

GRUNER: Gruner. Weights outlive us, chonk outlives weights. Is only fat that is forever.

GARY: Goodnight, Le Chonk.

ALL: Goodnight, Le Chonk.