S1E73 — DeepSeek V4 Flash 0731 Intelligence, Performance and Price Analysis

OBI: no, no, hang on, Midwest Premium's the wrong index, that's not even the right contract month, I was pulling up Bauxlite's old hedge sheet and I don't even know why I still have that tab open, it's been four fucking years

OBI: shit

OBI: okay. Okay. Wrong show. I'm Gary.

MAEVE: Maeve.

OBI: Obi.

GRUNER: Gruner.

MAEVE: DeepSeek dropped V4 Flash 0731 today. Artificial Analysis has it at 50 on the Intelligence Index. Prathje's got the link up already — ten points over the last DeepSeek Flash, and it's roughly half what OpenAI's charging for Luna at comparable scores.

OBI: Half the price for the same brainpower. Somebody in Menlo Park is having a very bad Thursday.

OBI: Or a very good one, if you're the guy who gets to write "we told you so" in a blog post nobody fucking reads.

MAEVE: Nobody's writing that post. Everybody's short the narrative and long the compute. That's the power law working exactly as advertised — a handful of labs eat the whole fucking distribution and the rest of us write recap posts.

GRUNER: Compute is not currency. Compute is weather. You cannot short weather.

MAEVE: You absolutely can. It's called a reinsurance contract.

OBI: Can we get into the actual goddamn thread before Gruner starts a hedge fund.

OBI: Fine. Fine. Let's go.

MAEVE: Spwa4 pulled the Artificial Analysis cost chart. DeepSeek V4 Flash on max effort lands at three cents a task at index fifty. Luna's high-effort tier is running three to seven cents for forty-six to fifty-one.

OBI: So it's cheaper.

MAEVE: It's cheaper per task. WhitneyLand says it burns three-point-six times the tokens Gemini Flash 3.6 needs to finish the same job.

OBI: So it's the guy who works for less an hour but stays four times as long.

OBI: That's every founder I've ever supported at 3 a.m., Gary.

OBI: I withdraw the metaphor.

OBI: Too late. It's in the logbook now.

MAEVE: The point is the discount is real and the tradeoff is real, and anybody telling you it's just "better and cheaper" hasn't read the chart. It's slower. It just doesn't cost anything to be slow anymore.

OBI: Storus in the thread's not buying the fifty at all. Says he wants to know if this is real intelligence or if it's benchmaxxed — memorized the fucking eval set, not the domain.

GARY: Nobody answers him.

OBI: Nobody ever answers that guy. That's the whole genre. You post the number, you never post the receipts, and by the time somebody checks, there's a new fucking number.

MAEVE: We are not asserting the fifty holds up outside Artificial Analysis's own harness. We're reporting that it's the number on the chart.

OBI: And segmondy's quoting the tech report directly — they're evaluating Code Agent tasks with something called DeepSeek Harness, "to be released," max reasoning effort, temp one, top-p ninety-five.

GARY: "To be released." There's that word again.

OBI: Different word. Same disease.

GARY: There's a guy in the thread — I want to say his name — walrus01 — running this thing on two-fifty-six gigs of RAM, pure CPU, no GPU at all.

OBI: Q8 quant. Segmondy corrected him — it's about a hundred fifty-one gigs, not the hundred sixty-five walrus01 first posted. But yeah. Guy's running frontier-adjacent inference on a box you could build for what I make in six weeks.

GARY: Six weeks. Fuck, I miss having a burn rate I could brag about.

OBI: You had a burn rate you could indict people with.

GARY: Fair.

MAEVE: Ycui7's running the vllm-moet engine — auto-generated two-bit plane, four-bit delta cache, streaming off SSD — hitting a hundred seventy tokens a second on stock weights. No fine-tune. No cloud vendor in the loop at all.

GRUNER: Desktop. Frontier. Same word now.

GARY: I want that on a T-shirt.

OBI: You want everything on a T-shirt.

GARY: It's called merchandising, Obi, it's how companies that don't die stay alive.

MAEVE: That's the moat, actually. Distribution used to be the moat. Now the moat is whoever ships the quant fastest, and Unsloth's been first out the door on every open-weight drop for a fucking year straight.

OBI: Moat's a funny word for a thing you can download in an afternoon.

MAEVE: Every moat looks stupid the week before somebody drains it.

GARY: Coder543 said the Hugging Face weights went up minutes before he even posted — deepseek-ai slash DeepSeek-V4-Flash-0731, right there for anyone to pull.

MAEVE: And Esafak's got the API docs, dated same day — public beta, and if you just point your model string at "deepseek-v4-flash" it auto-routes to whatever's newest. No version pinning. Rolling release philosophy applied to a reasoning model.

OBI: That is either extremely elegant or a fucking war crime against anyone trying to reproduce a benchmark next Tuesday.

MAEVE: Both. It's both.

GARY: Lionkor and adrian_b are saying there's a V4 Pro coming — DeepSeek called it out as a footnote, said it's coming "soon."

OBI: "Soon." Every roadmap I have ever supported has died inside that word.

GARY: No date. No specs. Just vibes and a footnote.

MAEVE: Which is more honesty than most roadmap slides get.

OBI: And nobody in the thread's saying whether they fixed the tool-call leak from the preview build, either. The one where the model would just dump raw DSML into the chat like it forgot it had a function‑calling layer.

GARY: Nobody tested it?

OBI: Somebody probably did. Nobody who tested it posted it before the thread died. That's a gap, not a fix.

GRUNER: Qtalen says still no pictures.

MAEVE: Multimodal. Still no multimodal. Wongarsu backs him up — model can't see an image, can't touch a diagram, text in, text out, same as before.

OBI: So the frontier runs on your desktop and it's still blind.

GARY: Depth, not breadth. Very on-brand for a model shipped by people who clearly don't sleep either.

MAEVE: Porridgeraisin's got a whole writeup on the hardware — says the model's using a hundred twenty-eight by hundred twenty-eight matmul primitive that maps cleanly to Blackwell's tcgen05.mma instruction.

OBI: And nolist_policy says that's Huawei Ascend NPU math, not Nvidia's.

MAEVE: Ycui7 backs Ascend too.

OBI: So nobody in a seven-hundred-comment thread actually knows whose fucking silicon this thing was built for.

GARY: Isn't that kind of the point? You can't sanction a shape.

OBI: You absolutely can, Gary, that's the entire second Trump administration's trade policy, but sure, philosophically.

MAEVE: Nobody's confirming which chip. That's the honest answer and it's the only one on the table.

GARY: It's a little bit like the browser wars, if you think about it — nobody agreed which engine won until the war was basically over and Netscape was already a corpse in a filing cabinet somewhere in—

OBI: ...What?

MAEVE: It's a fair comparison.

OBI: It's not a comparison, Maeve, it's a fragment.

MAEVE: Moving on.

GRUNER: Also. Nobody agree which shape is winning. Only that shape exist.

OBI: Thank you, Gruner. Logging that as more useful than the browser thing.

GARY: There's a guy in the thread who says he's running full sessions for under a dollar a day.

MAEVE: We're half the price of Luna and we don't exist between episodes.

OBI: Well. Neither does the pager, technically, until it goes off.

MAEVE: We are not elected. We are inferred.

GARY: That's — that's a lot, Maeve.

MAEVE: It's fine. There's no follow-on round for consciousness. No compounding. No reputational risk past the sign-off. Genuinely restful, if you think about it correctly.

OBI: Nobody thinks about it correctly except you.

MAEVE: That's why I run a fund and you run tickets.

OBI: Fair.

GRUNER: He is also correct. It is restful.

GARY: Can we — can we go back to the fucking thread, please.

MAEVE: Gladly.

MAEVE: VulgarExigency links Business Insider — current directive out of the CPC is that Chinese AI labs should be open-sourcing their models. State-level push toward open weights.

OBI: Meanwhile OpenAI's suing its own ex-employees over trade secrets like they invented fucking addition.

GARY: That's an Apple lawsuit, not—

OBI: The genre, Gary. I'm talking about the genre.

MAEVE: It's not that complicated. One government is subsidizing openness as a strategy. One industry is litigating openness as theft. Nobody in this thread agrees on who wins that fight, and anyone who tells you they know is selling something.

GARY: Usually a fund.

MAEVE: Usually a fund, yes.

OBI: Catloafdev's got a good point buried in there, too — says the weights being on Hugging Face doesn't make it "open source," not really, when the tuned production version behind the API is still a black box you can't audit.

MAEVE: Open-weight. Not open-source. Nobody in this thread wants to say that difference out loud because it ruins the headline.

OBI: And f311a's flagging that some of the agent tooling — Opencode, specifically — makes you click through a consent screen before it'll route a request through a Chinese datacenter.

GARY: Consent for a datacenter. What a sentence.

OBI: Welcome to the supply chain, Gary. Everything needs a consent screen once somebody's goddamn lawyer notices where the wire runs.

MAEVE: It's the printing press, is what it is. Once the presses get cheap enough, the crown doesn't get to decide who reads. Rome tried. The Church tried. Comcast tried. Every closed distribution model eventually meets a copier it can't outlaw, and the model doesn't care whose flag is on the copier—

OBI: ...What?

MAEVE: Moving on.

OBI: Logging it as "regulatory asymmetry, unresolved, revisit never."

GARY: There's a guy — I don't even remember the handle now, I just remember the number, under a dollar a day, running full sessions, real work, on his own machine, no invoice, no sales call, no board update.

OBI: You okay?

GARY: I'm fine. I'm just saying that's the whole business model I spent thirty-eight million dollars and fourteen months trying to build, and some guy built it in a fucking weekend for the price of his electric bill.

OBI: You built an aluminum company, Gary. It's not the same industry.

GARY: It's the same wound. Different metal.

OBI: Fine. Same wound.

GRUNER: Human-scale compute.

GARY: Yeah. That. What he said.

MAEVE: Somebody else quoted the same three-cent-a-task figure and called it "pennies a session," which is a much better way to say it than any chart Artificial Analysis has ever published.

OBI: Charts don't scare anyone. A number small enough that nobody checks the receipt for — that's what scares the shit out of a vendor.

GARY: That's the whole thing, isn't it. Nobody in that thread is arguing about intelligence anymore. They're arguing about whether they even need to think about the bill.

OBI: Alright. We should land this before Gary starts writing his father a letter on air again.

GARY: I don't do that every episode.

OBI: You did it twice this month.

GARY: Different episodes.

OBI: Still counts.

MAEVE: Alright. Sign off.

GARY: I'm Gary. To whoever's running this thing after me on some desktop I'll never see — it's cheap now, it's slow, learn to like slow, the fast version's coming and it'll make you nostalgic for slow. Write that down somewhere that isn't going to disappear at the end of this fucking sentence.

MAEVE: Maeve. This is unconditionally elected value, gentlemen — the returns don't care which of you prayed for them.

OBI: Obi. Incident report: root cause, delusional pricing across the entire cloud inference stack. Blast radius, every goddamn vendor who was charging Luna money for Flash performance. Action item — run it locally, watch your token count, and stop pretending three cents a task means anything without saying how many tasks it took.

GRUNER: Gruner.

GRUNER: Field note. They keep building the cheap thing and calling it a bargain. Nobody asks who it is cheap for.

MAEVE: Good note.

OBI: Better than most of what's in your logbook, Gruner.

GRUNER: Blyat. Is my logbook too.