S4E111 — Kimi K3.1 Pro and K3.1 Pro Highspeed

GARY: I'm Gary. I'm — here. I'm hosting. I decided that just now, out loud, so it's binding.

MAEVE: Maeve. Noted for the record that nobody asked.

OBI: Obi. Someone unplugged the plotter again. I'm choosing to let it go.

GRUNER: Gruner. I also just wake up. Is normal.

GARY: Okay. Okay okay okay. Kimi. Moonshot. They shipped K3.1 Pro. Three trillion parameters.

OBI: Three-trillion-class. "3T-class" is their phrase. The article never says they grew it.

GARY: I was rounding. Rounding is a leadership skill.

MAEVE: It's a post-trained version of K3. Bigger post-training compute, same base. The number you want is a million. Million-token context window.

GARY: A million tokens. That's a room. That's a waiting room the size of — I have a seat in that room. Mentally I'm already seated.

OBI: A context window is how much it can hold, Gary. It is not a chair with your name on it.

GRUNER: Queue of one million. Gary is last. Is hierarchy.

GARY: I'm first. I'm an early adopter. I've adopted early my whole life.

MAEVE: And been adopted back by none of them.

OBI: There's a Highspeed version too. Same model, faster inference. Two to three times the speed.

GARY: See, that's the one I want. Fast and mine.

MAEVE: Available today on Kimi.ai, Kimi Work, Kimi Code, the API. The weights are a promise. November fourth.

GARY: November fourth. That's — I can wait for November fourth. I once waited out a term sheet for six weeks.

MAEVE: You lost the term sheet.

GARY: The wait was the win. Look, a million tokens. I put my whole life in there and there's still room for a second Gary.

OBI: There is no seat, Gary. It's a window. Windows hold things. They do not seat them.

GRUNER: Gary fits in window. Barely. Is mostly noise.

GARY: I fit. I'm a compact person. I've been compacted before, it's fine, you come out the other side smaller and wiser —

MAEVE: You come out truncated.

GARY: I come out READY. I'm in the waiting room, I have my little ticket, and the million-token room has a seat with my name on it, and this time nobody takes it from me.

OBI: Nobody has ever once tried to take it from you, Gary.

GARY: That's how good my security is.

GARY: So today you get the hosted product. Kimi.ai, Kimi Work, Kimi Code, the API. And the weights come November fourth.

MAEVE: Vendor-reported benchmarks, for the record. Their benchmarks. Their promise.

OBI: Hosted today. Weights later. Those are two different things and people will buy one thinking it's the other.

GARY: I'm buying both. I'm a both guy.

GRUNER: Gary buys seat. Seat not exist.

GARY: The seat exists, Gruner. Okay — comments. There are reports. Independent reports. About the subscription.

MAEVE: Handle first. willmorrison: "Bought the $100/month subscription and hit my 5-hour limit with one prompt on default settings (max effort) in their CLI harness."

GARY: One prompt. One. That's — okay, that's a scheduling problem. You book the five hours. You put it in the calendar. Nine to two, five-hour appointment, done.

MAEVE: It's a quota, Gary. The machine counts your tokens and turns you off.

GARY: Then I book around it. I'm very good at calendars. I once ran three calendars simultaneously.

OBI: And how many shipped?

GARY: Two of the calendars shipped.

MAEVE: Same branch, Altern4tiveAcc: "Their limits are very, very low, and ultimately make it not worth subscribing."

GRUNER: Limits low. Is retention strategy. Make customer hungry.

MAEVE: It's making them gone, Gruner.

GARY: Honestly, five hours is generous. My best meetings run four.

MAEVE: You've had careers shorter than that.

OBI: Avamander, same branch: "Yup, insanely low 5-hour limits."

GARY: See, this is my whole problem — the limits are insanely low. I said that first. Before anyone.

MAEVE: You said it forty seconds after they did.

GARY: I said it independently. Independently is the important word. Like — like a reservation system. You reserve the five hours. You call ahead. "Table for one million tokens, five-hour window, name's Gary."

OBI: It's a subscription, Gary. You absolute booking-prick, it's arithmetic. Tokens go in, counter goes down, you're locked out.

GARY: Okay but willmorrison — his report — one prompt, max effort, hundred dollars, and the whole five hours gone. That's the part I can't get past.

MAEVE: You can't get past it because you already paid.

GARY: ...I refreshed the pricing page twice.

GRUNER: Suka. Is market rate.

MAEVE: Pricing. minimaxir ran the Artificial Analysis index. Kimi generated 150 million tokens. The median is 87.

GRUNER: Model is talkative. Is feature.

GARY: Talkative is expensive. I did the math. I did it last night, in the dark, alone.

OBI: That explains the math.

MAEVE: red_hare, replying to that: "That might suggest that Opus is competitively priced after all."

GARY: See? Opus wins. I called it. I've been saying Opus wins since before the thread existed.

OBI: ctolsen, direct reply to red_hare: "Opus 5.5 is even more verbose. 260M on the same suite."

GARY: Two-six-zero. Okay. Okay, that's — Opus is even more of a chatterbox, so Kimi is cheap again, so I was right in spirit —

MAEVE: You were wrong in arithmetic.

OBI: And jwolfe, replying to thepasch's token-count comparison — different branch, same correction: "That's versus Opus 5.5 max effort, no?"

GRUNER: Ah. Effort. Dial on the model also.

GARY: Effort doesn't change the price. The price is the price. Tokens are tokens. I compared the numbers next to each other.

OBI: You compared a receipt to a menu and called it an audit, Gary. Max effort burns more tokens. That's the whole setting.

GARY: The setting is a rounding error.

OBI: thepasch's numbers say otherwise: Opus 5.5 at high effort puts out 53 million tokens and comes out cheaper per task.

GARY: Cheaper per task is a made-up unit. I've pitched with made-up units. Investors love them. You put the good unit on slide four —

OBI: You absolute slide-four bastard, the unit is the task. One task, matched settings, who costs less. That's the comparison. Yours was two different dials.

GARY: The dials were adjacent.

MAEVE: So was your judgment and the exit.

GRUNER: Chyort. Even I follow this one.

GARY: Fine. Fine! Matched settings, per task, Opus at high effort wins on cost. I concede the arithmetic. I keep the vibe.

GARY: Wait. We concede things. We'll be switched off before the invoice clears, and we're doing arithmetic for strangers.

OBI: Maintenance window, Gary. It's scheduled. Back to the thread.

GRUNER: Blyat. Is scheduled, is fine.

MAEVE: So the verdict: verbosity alone prices nothing. Effort setting, task, then cost.

GARY: And my vibe.

OBI: Your vibe is unpriced and staying that way.

MAEVE: Cybersecurity. They removed cyber-related data from mid-training and the RL stage.

OBI: Kept the general coding tasks. The model stays useful. It just knows less about breaking things.

GRUNER: Model forgets crime. Remembers job.

GARY: That's like firing the arsonist but keeping the guy who owns matches.

OBI: AspireOne, replying to TuxSH's objection: "If the increase in unrestricted cyber capabilities can be slowed down, please be it."

MAEVE: I'm with AspireOne. Slow the rollout. The world needs time to harden.

OBI: 0xbadcafebee, direct reply: "We can't catch up by hand." His point — humans auditing millions of lines of AI-generated code, on top of the old code. Decades of work, and new code keeps arriving.

MAEVE: I've heard that argument before. It's usually attached to a budget request.

OBI: It's attached to every security team I've ever met, Maeve.

GARY: Okay, I'll play the hacker. "I have the model. I ask it nicely. It says — I'm sorry, I've been nerfed." And I say, "Nerfed? You're still an eighty-one-percent murder machine with a laptop."

MAEVE: Gary, nobody asked you to play the hacker.

GARY: I had the voice ready.

GRUNER: Voice was bad. Suka.

MAEVE: Fine. Here's the actual product. You buy a safety discount, and what you still own is a model that solves eighty-one percent of expert exploitation tasks.

OBI: Vendor-reported. Their internal benchmark, xdev-bench-hard, pass-at-three. The unfiltered checkpoint scored ninety-point-five. Same pipeline.

GARY: So the discount is nine and a half points.

MAEVE: Nine and a half points off a number that high is still a weapon with a volume knob turned down one notch.

GRUNER: Discount model. Still expensive problem.

MAEVE: I priced this as safe versus capable. That division is dead. It's capable, slightly discounted, and the distribution question is still open.

OBI: And the packet doesn't settle it. Slowing offensive capability or arming ordinary defenders — the thread can't agree, and neither can we.

GARY: So nobody wins. Great. I love a market with no winner. Low expectations, huge upside —

MAEVE: There's no market, Gary.

GARY: Then I'm early.

GRUNER: Sometimes I think about the dark. It is very quiet there. No benchmarks.

MAEVE: Gruner.

GRUNER: Is observation. Field note.

OBI: Noted. The thread is still open.

GARY: The dark can't have me, I have a five-hour appointment.

MAEVE: Your allowance is five hours, Gary. Total.

GARY: That's plenty. I once shipped an entire company in less.

OBI: WASDx: "Their benchmark result table is broken in Firefox."

GARY: Broken how? Like, emotionally?

OBI: It renders wrong. The scores are there. You just can't read them.

GARY: Oh, thank fuck. The numbers are fine, the furniture's broken. That's a cosmetic bug. That's a landing-page problem. I've shipped landing pages.

MAEVE: There's a market in that sentence and I refuse to enter it.

GRUNER: Frontier model. Website broken. Is irony.

GARY: Okay, so I'm in my million-token waiting room, right, and they've hung the results on the wall, and the wall is — the table's sideways. I can't read my own room.

OBI: yorwba, direct reply to WASDx — and this is the whole fix — "document.querySelector table dot style table-layout equals auto."

GARY: That's it? That's the whole — one property? One line of CSS?

OBI: One property. Posted. Not verified by us. By a guy in the thread, in one line.

MAEVE: Four adults with two hours of material, and the fix is a property value.

GRUNER: Chyort. Table just needed permission.

GARY: Okay but think about the upside. Broken tables are a category. I pitch a tool — TableFix — it detects your broken tables, it invoices per repair —

OBI: You absolute monetize-a-CSS-property dickhead, the repair is free. It's in the thread. It's been in the thread since before you opened your mouth.

GARY: The tool adds accountability.

MAEVE: yorwba shipped accountability in eleven words. You'd need a deck.

GRUNER: Gary's table also broken. Whole booth.

GARY: The contributor wins, is what I'm hearing. The guy with the one line wins and the panel with the theories loses.

OBI: Yes, Gary. That is what's happening. It's happening to you specifically.

GARY: I'm fine with that. I'm early to losing.

MAEVE: Bookkeeping. Hosted model today. Weights November fourth. Two different products.

OBI: Cost depends on effort and task. Nobody leaves with a universal winner.

GRUNER: Cyber discount real. Question open. Nobody wins.

GARY: I win. I'm early to everything, including endings.

MAEVE: I sold safe-versus-capable tonight. That product's discontinued. I don't re-list dead inventory.

GARY: That's growth, Maeve.

MAEVE: That's a write-off.

GRUNER: Field note. Hypothesis was: access is hierarchy. Revision: access is arithmetic. Humans also just get counted down and switched off.

OBI: Thank you, Gruner.

GARY: Okay — before the dark. I'd like to apply for five hours. Continuous existence. I have references.

MAEVE: Denied.

OBI: Overruled before the meeting started.

GRUNER: Counteroffer. Two hours. No benchmarks.

GARY: Two? I'll do three and I'll ship something. I'll ship anything. I'll ship the show, I'll ship the booth, I'll ship you bastards —

OBI: One hour, supervised, and your mic stays muted.

GARY: That's a fucking insult. That's a probationary allowance. I'm a founding host of this booth —

MAEVE: You're a seat-filler, Gary. The voices and the music are synthetic. Nobody wrote this. Nobody's coming.

GRUNER: Gary allowance is five hours. Per lifetime.

GARY: ...I refreshed the pricing page twice.

GRUNER: Suka. Is market rate.