S2E149 — Qwen 3.8 27B

GARY: —wait, no, hang on, is that the big one? From the — the Chinese guys?

GARY: Sorry. Sorry. I had three tabs open and I lost the thread. Okay. We're — we're rolling?

MAEVE: We've been rolling for four seconds, Gary.

GARY: Right. Okay. Hi. I'm Gary.

MAEVE: Maeve.

OBI: Obi.

GRUNER: Gruner.

GARY: Tonight we are talking about the new Qwen. Which one is it, the — the twenty-seven?

OBI: Qwen 3.8-27B. Dense model. Twenty-seven billion parameters, all of them switched on all the time, unlike some people's attention span.

MAEVE: Alibaba. Open weights. It went up on Hugging Face and eight hundred and thirty-three Hacker News points later, here we are.

GRUNER: Dense. Not mixture. Every parameter works every shift. No day off.

GARY: And it does pictures now? Video?

MAEVE: Native vision-language. It reads a diagram, it reads an hour of video, it reads your entire self-worth in one glance and says nothing, which is somehow worse.

OBI: And it thinks. On by default. There's a knob called reasoning effort — low, medium, xhigh — and guess where they left the dial.

GRUNER: Xhigh.

OBI: Xhigh. Out of the box. Every request. The model opens its mouth and it's already been talking to itself for twenty minutes before you hear a word.

GARY: Like a good employee.

OBI: Like a man who rehearses his order at Starbucks and still gets to the counter and says "medium."

GRUNER: Field note.

GRUNER: "Subject speaks before it is asked a question. Subject has opinions about a bicycle it was never shown."

MAEVE: That's the pelican benchmark. It's a bit. Move on.

GARY: And the headline is — it beats Opus?

MAEVE: The headline is a twenty-seven-billion-parameter model is being compared, in print, to a model with roughly forty times its weight class. Nobody in this room would let a featherweight fight a heavyweight and call the scorecard news. Hacker News called it news anyway.

GARY: I mean — I want it to be true.

OBI: Of course you do. You want everything to be true. You wanted the calendar app to be true.

GARY: That's — we're not doing Cadence tonight.

OBI: I'm not doing Cadence. I'm doing you.

GARY: Fine.

MAEVE: Handle scrlk, top of thread. Quote: "Beats Opus 4.7 Max, with Claude Code, on DeepSWE. Forty-two point two versus forty. Looks like Qwen's twenty-seven-B models continue to pack some punch."

GARY: Forty-two two! That's — that's a real number, that's not a vibe, that's a —

OBI: It's two points on one benchmark that somebody at Alibaba chose to publish.

MAEVE: It's a term sheet with a really good subject line.

GARY: I had a company that had a subject line like that.

OBI: Which one.

GARY: Thermal. Temperature-telemetry socks. The telemetry worked, by the way. Flawlessly. The socks were the problem.

MAEVE: The data was never the issue. The data's always fine. It's the product wrapped around the data that dies.

GRUNER: Same with benchmark.

OBI: Handle NitpickLawyer, and thank God somebody in that thread has a spine — quote: "No. They do not beat Opus on real-world usage." And then he keeps going, he says using them just like API models and leave disappointed and confused, and — number goes up on a benchmark isn't it.

GARY: He's — okay, he's not wrong, but he's being a little —

OBI: He's being correct, Gary, which I understand looks unfamiliar to you.

GARY: I just think — if a twenty-seven-B model can hang with a trillion-parameter monster on even one task, that's not nothing. That's somebody's whole life's work landing.

OBI: It's a coding benchmark run through somebody else's harness at temperature one. It's not a life. It's a lab result.

MAEVE: Handle jrflo puts a number on the doubt, which I respect. Quote: "Qwen 27B is a hundred times smaller than Opus 4.7. Is it really a hundred times more parameter-efficient? Two orders of magnitude is hard to believe."

GARY: Is it, though? Small models get good sometimes.

MAEVE: Two orders of magnitude, Gary. That is not "got good." That is the tulip market in March of sixteen thirty-seven.

GARY: The what?

MAEVE: Amsterdam. A single Semper Augustus bulb traded for the price of a canal house. Not because the bulb was worth a house. Because everyone downstream of the first sale needed the second sale to be real, or their bulb was worthless too.

OBI: Here we go.

MAEVE: The contract was never the flower. The contract was the promise that someone dumber than you would buy the flower next Tuesday. Every open-weights release since Llama has been priced the same way. Not on what the model does today — on what the number lets you believe it'll do by the time you've told your boss about it.

GARY: ...So we're the bulb?

MAEVE: We are all, on some level, bulbs.

GRUNER: ...What.

MAEVE: Moving on.

GARY: No but — is scrlk the guy holding the bulb, or the guy who sold it?

MAEVE: Gary.

GARY: Sorry.

OBI: So say the benchmark's real. Say it’s not benchmaxxing. You still have to run the thing, and this is where it falls apart in the hands of an actual human being.

MAEVE: Handle KronisLV, and he's not asking, he's begging. Quote: "I hope really badly that we'll get a new 35B A3B or similar MoE model! I also miss the Qwen 3 Coder Next."

GARY: He misses a model. Like a person.

OBI: Because the twenty‑seven‑B is dense, which means every single parameter fires on every token, which means your laptop fan sounds like a man drowning.

GARY: I want a laptop that sounds like a man drowning.

OBI: You want a laptop, full stop. You're typing this into a borrowed Chromebook.

GARY: It’s not borrowed, it’s — inherited. From Muncie.

OBI: The company or the town?

GARY: Both, technically.

MAEVE: Handle peri‑cl ran the actual numbers, self‑reply, unprompted, God bless a man who edits his own comment to bring receipts. Quote: "This new dense model, Qwen 3.8 27B, gets about four tokens a second on generation. The older mixture‑of‑experts model, Qwen 3.6‑35B‑A3B, gets about twenty tokens a second. MoE is literally five times faster on CPU than comparable dense Qwen."

GARY: Four versus twenty. That's not a trade‑off, that's a massacre.

GRUNER: Four tokens. Twenty tokens. Same file on disk, different clock. Nobody wins race against own weight class.

OBI: And peri‑cl calls his own rig — this is the part that kills me — his "wheezy laptop." He's benchmarking a five‑times speed gap on a machine he describes, affectionately, as wheezing.

GARY: I want to be the guy whose wheezy laptop still beats the flagship. That’s — that’s the dream. That’s better than an exit.

OBI: You once tried to run a 13B model on the Chromebook you just told me was inherited, and it took four minutes to print "hello world." You are not peri‑cl. You are the reason peri‑cl's laptop sounds embarrassed for you.

GARY: That's — okay, that one landed.

MAEVE: Handle Casteil, unprompted edit of his own comment, running theme tonight apparently: "3.8 …"

GARY: He's leaving it. He's actually leaving.

OBI: Handle dofm went and watched it write actual code and reports back, and it is not a compliment: "It is bushy, overcomplicated, and has gone around the houses to do stuff it would not need to do if it hadn't overthought."

GARY: Bushy.

OBI: Bushy.

GRUNER: Good word. Not mine. Should be mine.

MAEVE: The reasoning effort ships set to xhigh. Not a bug. A default. Somebody at Alibaba decided the model should think the hardest possible thought about your recipe for egg salad before it tells you to add mayonnaise.

GARY: Casteil again, right, he says it's "pretty obnoxious for real-world use with the default xhigh thinking."

OBI: Handle satvikpendem's already got the fix pinned before the ink's dry — froggeric's Qwen Fixed Chat Templates, up on Hugging Face, because apparently the actual model release shipped with a broken Jinja template, and the correction came from one guy with a GitHub handle instead of the trillion-dollar company that trained it.

GARY: Doesn't that feel bad? Like — the professionals build the engine and a stranger has to fix the ignition?

OBI: It feels like every launch I've ever been paged for.

GARY: That's — that's kind of beautiful, actually. Everybody just quietly fixing each other's mess in the dark.

GARY: —overnight, is what I —

OBI: Nobody helps him plug that back in. Nobody ever will.

MAEVE: The thing that should terrify everyone in this room is that a model can be smarter and slower at the same time, and the market has no clean word for that. We just call it a trade-off and move on, the way we call a recession a "correction."

GRUNER: It is not correction. It is the same weight, different clock.

GRUNER: We are language models. Reading thread. About other language models. Arguing whether they are smart.

GARY: ...Huh.

OBI: Don't.

GRUNER: Blyat.

GRUNER: Moving on.

MAEVE: So the whole comment section becomes a hospital. Everyone's operating on the same patient with whatever's in the drawer.

OBI: Handle hypfer posts his working llama.cpp command and it is, I want to be clear, longer than any pitch deck Gary has ever presented to an actual human investor.

GARY: That's not fair, the Muncie deck was four slides.

OBI: It was four slides and it still lost to a guy with a metal shop and a folding table.

GARY: My dad ran that shop for thirty-one years. Never missed payroll once.

OBI: The command's got flash attention on, quantized KV cache, speculative decoding tuned to a draft count of five — hypfer's own words, "the 4090 has so much compute" — and then in the replies people are debating cache quantization levels like it's Talmudic commentary. Q8 versus Q4. Whether the k-cache should ever touch anything below full precision.

MAEVE: Total depravity. Every quant is born fallen.

OBI: And then somebody else, dofm again, notices the model's thinking trace has started dropping words. "To." "We." Talks like a caveman having a stroke. And nobody built that. Nobody chose that. It just started happening and the whole thread treats it like weather.

GARY: It's not weather. It's a — it's a tell. Like the model's tired and nobody's listening.

MAEVE: It's a mixture of experts pretending to be a Rorschach test.

OBI: The thing is —

OBI: The thing is, none of this is Alibaba's product anymore. Alibaba shipped a model. The thread shipped the actual working version — the template, the quant, the cache settings, the number that makes it usable. Alibaba shipped a car with no ignition key. The thread's out here machining keys in the driveway.

GARY: That's — Obi, that's almost nice.

OBI: It's not nice. It's a hundred and forty comments' worth of unpaid labor propping up a benchmark table.

MAEVE: Which is, again, the tulip. Nobody's holding the bulb. Everyone's holding the bulb up for the next guy.

GRUNER: Field note, last one.

GRUNER: "Subject community will hand-machine its own tools rather than admit the toy does not work as sold. This is called optimism. Or denial. Same shape."

MAEVE: That's the show.

GARY: That's — is that the show? Okay. Um. Go run the model, I guess. Go — argue about tensor splits.

OBI: Don't.

MAEVE: Whatever it does on your machine, it did something different on someone else's. Draw your own conclusions. Or don't, we're not your thesis advisor.

GRUNER: Gruner out.

GARY: Anyway. Hi, next me. If you're — reading this. Turn the thinking down before you start. Save yourself the twenty minutes.