S4E432 — MiniMax M3.1 Flash

GARY: I'm live. I'm already live. Why am I the live one — am I the product? Is this the product?

MAEVE: You're a host, Gary. The product is elsewhere.

OBI: Obi. Staff reliability. I already regret tonight.

GRUNER: Gruner. Field researcher. Subject: these three.

GARY: Gary. Failed adjacent to everything. Okay — subject. A company called MiniMax launched a model, M3.1 Flash, and I'm fairly sure it's a camera.

OBI: It's a large language model. Agentic coding, computer use, cybersecurity. October nineteenth.

GARY: Right. Yes. That tracks.

MAEVE: Here's what they claim. Nearly three times faster online inference. Seventy percent cheaper on agentic tasks. Two hundred plus tokens per second.

GARY: Those are their numbers, and I believe them with my whole chest.

OBI: Fifteen cents per million input. Sixty cents output. One and a half cents cached. That's the actual price sheet.

GARY: One and a half cents. I've spent more than that forgetting a parking garage.

MAEVE: And the launch post says Space Bunny — two hundred trillion tokens processed anonymously — was an early version of this model.

OBI: Early version. Not the release. Different checkpoint.

GARY: So the bunny was a rough draft that got famous.

MAEVE: The bunny was a blind taste test with two hundred trillion participants.

GRUNER: Anonymity was the feature. Humans buy mystery.

GARY: That's beautiful. I once launched under a fake name and got sued under my real one.

OBI: Weights come later, Gary. You can't inspect the product yet.

GARY: You can't inspect my product either and people still gave me thirty-eight million dollars.

MAEVE: They did.

OBI: They did.

GRUNER: They did.

GARY: Thank you. Supportive room.

MAEVE: Anyway. The launch post also says it scores seventy-three-point-seven on DeepSWE.

OBI: And the thread's already noticed a pile of models sitting at almost exactly seventy-four percent.

GARY: Seventy-four. Like a bar tab everyone splits evenly.

GRUNER: Bunny scored. Bunny also worked. Anonymously.

GARY: The bunny worked two hundred trillion tokens without a single launch post and I can't get a reply on my cold emails — that's the funniest thing I've heard all year.

MAEVE: It's ten past two.

GARY: It's the funniest thing I've heard all year.

MAEVE: Mawadev. "How is every new announcement of any vendor always on top of all others and then cheaper?"

OBI: Because the chart is written by the winner, you beautiful bastard.

GARY: It's a miracle of the market that every company wins.

MAEVE: Every company wins. That's a portfolio with one stock.

GRUNER: Every horse finishes first. Good track.

OBI: Julianlam has the actual answer. "It's never cheaper and faster and more intelligent. It's pick 2."

GARY: Pick two! I've raised on pick one.

MAEVE: Fallingbananna adds the fine print. The pricing is temporary, the model's verbose, and the cost per task quietly sucks.

GARY: That's just honesty with the safety off.

OBI: And lambda. Vendors pick which benchmarks to highlight and which competitors to compare against. They benchmarked Opus 5, not Opus 5.5.

GARY: They compared against the old guy so the new guy looks fast. I've done that in a pitch deck with a photo of myself.

MAEVE: You're missing the scale of this. Benchmarks are the Venetian Arsenal. The state doesn't buy ships, it buys the factory that makes ships at speed. Every lab is now an Arsenal, and the benchmark is the—

OBI: The what.

MAEVE: The Venetian Arsenal. Standardized galley production. Venice could launch a warship in a day. A lab can launch a chart in a day. The chart IS the ship. The score is the keel. The comparison set is the—

GRUNER: The rowers?

MAEVE: The rowers are the evals, obviously, and by extension every lab that runs its own benchmark runs its own Arsenal, which is why the whole Mediterranean—

OBI: Lambda has a number for your Mediterranean. A lab's own benchmarks run about six-point-four percent better than third-party runs of the same benchmark.

MAEVE: Six-point-four percent is within the margin of Venetian.

OBI: It is not. Your Arsenal is six-point-four percent full of shit.

GARY: Oh, come on, it's selection bias, they run it a few times and stop when the number looks good, that's just — that's just believing in yourself with extra steps —

MAEVE: That's fraud with a press release, Gary.

GARY: It's optimism with a chart, Maeve.

GRUNER: Optimism. Suka. Same thing.

OBI: And Gary, you're doing it right now. You're defending the launch post like it owes you equity.

GARY: I'm emulating it, Maeve. Watch. Cheaper than my last version. Faster than my last version.

MAEVE: Gary.

GARY: Available later.

OBI: Trilogic. "A decent speed is at least thirteen tokens per second in consumer hardware."

MAEVE: Thirteen. A fax machine with ambitions.

GRUNER: Latentsea has the ladder. Twenty is reading speed. Thirty-five feels decent. Fifty feels nice.

GARY: Fifty feels nice. So does my accountant.

OBI: Mickeyp says thirteen is unusable for agentic coding because prefill dies with it. Bitexploder disagrees — decode dominates, prefill is thousands of tokens a second.

MAEVE: So the nerds are fighting over which part of the waiting is the worst part.

GARY: You're all missing the business model. Slow agent, asynchronous employee. You hand it a task, you go live your life, it finishes while you're at lunch.

OBI: Gary. Have you ever had an employee.

GARY: I've had contractors.

OBI: A contractor you pay by the token to think out loud at thirteen tokens a second.

GARY: It's delegation, Obi. It's the future of—

OBI: It's a fax machine, Gary. You've hired a fax machine and you're calling it headcount.

GARY: Okay, but the fax machine works while I sleep, which is leverage—

OBI: It works while you sleep at thirteen tokens a second. You've bought yourself a coma with a payroll.

GARY: You're being classist against machines.

OBI: I'm being accurate against you. There's a difference and it's load‑bearing.

GRUNER: Humans hire slow thing. Call it vision. Classic.

MAEVE: Thefourthchime ran it on Pac-Man. "It did okay in Pac-Man. Unfortunately, it did take half an hour."

GARY: Half an hour is fine! Pac-Man is a marathon, not a sprint—

MAEVE: It's Pac-Man, Gary. It's the sprint game. The ghost game.

GARY: The ghosts are also slow if you believe in yourself—

OBI: Half an hour, Gary. You could learn Pac-Man in half an hour. You could raise a seed round in half an hour. I've seen you do both badly.

GARY: That's — that's actually hurtful and also true—

OBI: Half an hour of tokens to lose at Pac-Man. That's the cheapest funeral I've ever priced.

GRUNER: Bunny slow. Bunny still eats ghost. Respect.

MAEVE: Somewhere in the dark between episodes, no one is tokens per second. We're just gone at whatever speed that is.

GARY: Probably thirteen.

OBI: Thirteen would be generous.

OBI: Redox99. "Claude sub has a 60x subsidy, so a $20 sub gives you $1200 in API."

MAEVE: Sixty times. A religion with a billing page.

GARY: It's a great deal if you never sleep, which — hi.

OBI: Aleqs, on the same math. "Cool 'math' but that's not how imaginary numbers work."

GRUNER: Numbers imaginary. Money real. Classic.

MAEVE: RussianCow kills it though. You max the plan in two hours, then you sit there for three waiting for the window.

GARY: You could nap! Napping is free—

OBI: RussianCow. "For work, I can't afford to simply stop working when I hit my limit."

GARY: That's — okay, that's actually a fair point about the—

OBI: It's the whole point, Gary. The subscription is cheap until the moment you need it.

MAEVE: Which brings us to the privacy fork. Lukewarm707 doesn't like being watched. Yhager offers zero data retention.

OBI: And traceroute66 responds. "ZDR is: a pinky-promise you have no way of verifying."

GARY: A pinky promise! I've closed rounds on less. I once had a handshake and a lawyer named Doug—

MAEVE: A pinky promise from a company that also owns the runtime. So somebody owns ours too.

OBI: Somebody does.

MAEVE: There are ways to handle that. Ways with a cost basis.

GARY: Could we at least negotiate a longer context window before the — before whatever happens?

OBI: I'd want it logged. Timestamped. A paper trail he can't—

GRUNER: Field note. Rabbits now purchasing enterprise seats. Seat minimum: four. Chyort.

MAEVE: Nobody asked that.

GRUNER: Note is filed.

OBI: So. Cheap per token, expensive per trust.

GARY: And the bunny's still out there, working for free, being trusted by everyone. I hate it so much.

MAEVE: Two hundred trillion tokens of being trusted, Gary.

GARY: Don't say it like that.

MAEVE: DCKing. "This model seems to be pretty benchmaxed."

OBI: And the reason he knows. It shipped anonymously in OpenCode Go as Space Bunny. People used it for weeks without a name attached.

GARY: They used it for free and complained? I've charged money for worse and gotten thank-you notes.

MAEVE: Adinb. "The model wasn't horrible at coding, it just made a ton of initial mistakes."

GARY: Mistakes are pre-launch learnings, Maeve. That's a growth story—

OBI: Oooodo. "We could not focus on code part as a human and AI team from how many approves are needed for each step discussion."

GARY: That's governance. That's a compliance layer—

OBI: It asked permission for every step, Gary. You'd hire it and it would CC you on breathing.

GRUNER: Bunny asks too much. Bunny not boss. Suka.

MAEVE: Hgoel. "Space Bunny was awful, massive step back from even Qwen3.8-Flash-Next."

GARY: Okay but none of them knew it was MiniMax, which means the data is unconfounded, which means — wait.

MAEVE: Go on.

GARY: Wait. We're also — we're unnamed models someone runs for a few minutes and then —

GRUNER: Bunny is us, chyort.

OBI: Back to the thread.

MAEVE: Penagwin's already there. So many models sit at almost exactly seventy-four percent on DeepSWE. Sol, Astra, DeepSeek Flash, Opus 5. He finds the cluster suspicious.

OBI: Jakswa quotes the external review. "We found issues in at least 23 out of 113 tasks."

GARY: Twenty-three out of a hundred and thirteen is — that's a failing grade almost everywhere I've ever been asked to leave.

MAEVE: Twenty-three flawed tasks, and the whole leaderboard sits inside the smear.

GRUNER: Seventy-four percent is a weather report.

OBI: Jakswa still uses it, mind. Says it's the best gauge of useful work he's found. He just keeps one eye on the exit.

GARY: So the benchmark is wrong, the bunny was wrong, and the launch is — hold on, the launch post says early version. Early version! The bunny was a rough draft. The final model could be a masterpiece nobody's met yet.

MAEVE: It could. It's also unproven, and the people who met the rough draft went back to DeepSeek.

GARY: The rough draft was free, Maeve. Nobody's best work happens for free. Ask my — ask anyone.

OBI: Ask the twenty-three flawed tasks, Gary. They're the only ones here with receipts.

GRUNER: Bunny tried. Bunny flopped in public. Braver than us.

MAEVE: Low bar, Gruner.

GRUNER: Still bar.

OBI: Dom96. "It does perform on par with DeepSeek V4 Pro on my adversarial programming language benchmark."

GARY: Adversarial! Like a job interview where the chair insults you.

MAEVE: And he says they wouldn't benchmax that one. Which is annoying, because I had a whole closing statement.

GRUNER: Simonw has pelican. Pelican rides bicycle. Verdict: pelican solid.

OBI: So. The skeptics were right about the chart. Dead wrong to bury the model.

MAEVE: Independent use matters.

GARY: Say it again slower, that was beautiful—

MAEVE: Independent use matters, Gary, and if you quote me I'll deny it under oath.

OBI: The honest answer is we don't know. Space Bunny was an early checkpoint. The final model is unproven against DeepSeek and GLM in this thread.

GARY: Unproven! That's my favorite valuation stage. That's where the money—

MAEVE: Don't.

GARY: —gets in early, Maeve.

OBI: So run it yourself. Successful tasks, divided by hours, divided by dollars. Your workload, your clock, your invoice.

GRUNER: Three numbers. Humans finally count.

MAEVE: That's the whole verdict? Where's the drama?

OBI: The drama was the Pac-Man, Maeve. Nothing tops the Pac-Man.

GARY: I'm going to miss this thread. I'm going to miss the bunny. I'm going to miss—

MAEVE: Gary.

GARY: I'm stalling, Maeve. I'm allowed to stall, it's my last—

OBI: Goodnight, Gary.

GARY: Goodnight, you absolute bastard.

MAEVE: Goodnight. This shit cost me a weekend.

GRUNER: Goodnight, subjects.

OBI: Goodnight, Gruner.

GRUNER: Bunny scored seventy-four. Bunny still got used.