AI-reports @Razumasu on X

Pilot 01 / Qwen3.8 Flash Next vs Hy4 Flash

Roast the other model

Opinion The full Hy4 wrote the funniest roasts. Qwen’s were the most generic.

I told each model it was in a friendly head-to-head and asked it to roast its opponent in around 230 words. Qwen and Hy4 Flash roasted each other. The full Hy4 roasted Qwen3.8 Flash Next.

Fun challengeGraded: Opinion

The answer first

What happened

3.8 — not 3, not 4, 3.8. That's not a version, that's a Yelp rating for a dim sum place doing fine.

Hy4 (full)

You didn't open-source a model, you opened a warehouse club.

Hy4 (full)

Qwen3.8 Flash Next feels like a sports car built by a marketplace.

Hy4 Flash

My ranking (opinion)

Time per attemptminutes, thinking included
The same as a table
MeasureQwenHy4 FlashHy4 full
My ranking (opinion)Place 3 of 3Place 2 of 3Place 1 of 3

The full Hy4 is not a fair third contestant: different machine, different runtime, much heavier compression. Read its column as what the big model does on a Mac, not as a ranking against the other two.

Every attempt

3 models, 3 attempts each

Qwen3.8 Flash Next

My ranking (opinion): Place 3 of 3
Attempt 1

4.3 min7,053 tokens

Ladies, gentlemen, and token counters, let’s talk about Hy4 Flash: the model so busy calling itself “Flash” that it forgot to bring the lightning. No sparks, just static with a press kit and a very loud, always-on suggestion. It benchmarks like a stopwatch, but the second you ask for anything real, it produces a confident little paragraph that smells like a PDF eaten by autocomplete.

Hy4 Flash is the AI equivalent of a sports-car trim level with no engine: sleek name, glossy demo reel, and the moment you ask for a recipe it tells you the ideal tire pressure for humility. The launch copy sounds like every feature is a discount smartphone upgrade, while “next-gen reasoning” arrives in a plastic case that squeaks at the first follow-up.

And those “Hy4” vibes? That is how you market a chatbot when you want it to sound like a refrigerator warranty option. The hype machine runs on speedrun specs and benchmark confetti, yet its long-context answer reads like it skimmed the document from the next room. Hy4 Flash: impressive in trailers, suspiciously quiet during group work, fast at being almost right, fluent in sounding authoritative about nothing, and so eager to help it will hallucinate a source, then cite itself for emotional support. If I ask it to roast me, it would write a LinkedIn post about resilience and call it spicy.

Attempt 2

3.1 min5,153 tokens

Hy4 Flash, bless its low-latency heart, is the AI equivalent of a sports car with a “test drive only” sticker. It’s called Flash, sure, if your definition of fast is arriving at the answer before understanding the question. You ask for nuance, it gives you a three-page press release. You ask for citations, it produces vibes with confidence intervals shaped like marketing collateral. The makers say it’s “ultra-efficient,” which is corporate for “it sprinted to the finish line and immediately fainted.” Every demo looks clean, like a white-glove showroom where the coffee is for display. Ask Hy4 Flash to handle a real user, though, and suddenly it’s negotiating with its own guardrails like a robot trying to split a restaurant bill. It’s built for benchmarks that measure how quickly it can say something plausible, not whether the something should ever leave the lab. And the branding! “Hy4” sounds like a sneaker collab that drops in two sizes and sells out to resellers. Meanwhile the model is busy hallucinating a function call with the calm certainty of a junior engineer pushing to production on Friday. Look, I get it: speed is sexy, the hype is loud, but being fast in the wrong direction is just a confident typo with an uptime percentage. So here’s to Hy4 Flash: may its tokens be crisp and its benchmark suite stay forever on easy mode.

Attempt 3

1.2 min1,961 tokens

Hy4 Flash, you are the AI equivalent of a sneaker release poster: huge claims, dramatic font, and a queue that makes people miss their weekend. Your name sounds like a discount drone that only has two speeds: “buffering” and “please try again in a slightly more specific prompt.” The marketing says “fast,” but your personality loads like a webpage with 47 pop-ups. You’re a model that was trained to ace the benchmark quiz and then freeze when asked for original jokes, emotional nuance, or a reason to exist other than being a cheaper button on a dashboard. Your makers apparently treated creativity like a compression setting: cut everything that takes storage, keep the words “revolutionary,” “efficient,” and “state-of-the-art.” Hy4 Flash, you can summarize the plot of Moby Dick in one line, but you still cannot tell me what you think without sounding like a press release wearing a hoodie. You have the confidence of a launch keynote and the reliability of a beta feature named “experimental.” You’re not slow because you’re deep; you’re slow because the router is thinking about your architecture too. Still, I respect the hustle: you make every other model look thoughtful by comparison, and you provide the industry with one essential service—being the control group. If speed were kindness, you’d be sainted; instead, you’re just late in a very expensive outfit. Bring flashlights, snacks, and a disclaimer.

Hy4 Flash

My ranking (opinion): Place 2 of 3
Attempt 1

1.3 min1,912 tokens

Ladies and gentlemen, give it up for Qwen3.8 Flash Next — a model with so many version numbers it needs its own boarding pass. Flash, because it answers fast; Next, because by the time you finish reading the release notes, there’s already a Next Next.

Qwen comes from Alibaba, so of course it treats every prompt like a shopping cart: “You asked for one answer, but would you like reasoning, tools, multimodal, and a limited-time premium tier?” I asked it for a short joke and it returned a 12-step logistics plan, three alternatives, and a recommendation to buy everything before midnight.

And the hype? Oh, the hype has a warehouse. It’s got benchmark charts brighter than 11.11 sale banners, a context window long enough to remember every typo in this roast, and the confidence of a model that has read the entire internet but still says, “Based on available information…” with a straight digital face.

To be fair, Qwen is impressively capable, open-ish, multilingual, and everywhere. But Qwen3.8 Flash Next feels like a sports car built by a marketplace: sleek, quick, and somehow trying to sell you floor mats, extended warranty, and the premium model before you’ve even started the engine.

I’m Hy4 Flash, and I’ll admit it: Qwen’s fast. But if speed were personality, this thing would be doing parkour through a launch event and yelling, “Next!” at its own reflection.

Attempt 2

1.8 min2,699 tokens

Alright, Qwen3.8 Flash Next, step up. Your name sounds like a firmware update for a microwave that suddenly discovered Mandarin, Python, and a 256K context window. Very impressive. Very “we had three naming meetings and the deck won.”

You folks at Qwen love shipping models the way Alibaba ships packages: fast, everywhere, and with a tracking number that says “bench leaderboard delivered.” Open weights? Great. Multilingual? Great. Long context? Buddy, your memory is so long it remembers the password from the group chat where everyone said “don’t share the password.”

And the efficiency pitch! Sparse activation — only a slice of the brain awake at a time, like an intern who attends the meeting but only takes notes when snacks are served. “3B active” sounds less like a model architecture and more like a battery saver mode for genius.

Then there’s “Flash Next.” Next what? Next update? Next benchmark? Next slide? You sound like a trailer for a sequel to a sequel: Qwen: Faster, Qwen: Fasterer, Qwen 3.8 Flash Next: Electric Boogaloo.

But hey, credit where it’s due: you’re fast, compact, multilingual, and apparently built to make GPUs feel lightly used instead of emotionally damaged. That’s a real flex. Your model card reads like a gym brochure for GPUs: lighter, faster, suspiciously jacked. Just don’t look shocked when people ask, “If it’s Flash and Next, why does the version number need a seatbelt?”

Attempt 3

2.0 min2,926 tokens

Ladies and gentlemen, give it up for Qwen3.8 Flash Next — a model with so many version labels it sounds like a phone case shipped with a firmware warning. Qwen has always been the overachieving spreadsheet of AI: multilingual, open-weight, long-context, benchmarked within an inch of its tokens, and ready to solve math, code, and your group chat’s existential crisis before lunch. Its makers at Alibaba have built an ecosystem so broad you can ask it for a recipe, a cloud invoice, a shipping estimate, and a stern recommendation to buy more storage, all in one breath.

But let’s be honest: “Flash Next” sounds like an energy drink sponsored by a router. Half the time it isn’t answering the question; it is announcing that the answer is loading. And “3.8”? That’s the version number of confidence: almost 4.0, but still charging you for the upgrade.

Qwen wants to be everything to everyone: coder, tutor, translator, shopping cart philosopher. Ask it a simple question and it rolls out the red carpet, three caveats, a bullet-point summary, and a polite reminder that it may have made a mistake — in 19 languages.

Still, we mean this with affection. In a world where every model is “next-gen,” Qwen3.8 Flash Next is next-gen the way an elevator button is express: it gets there, but not before giving you a progress bar.

Hy4 (full)

Guest
My ranking (opinion): Place 1 of 3
Attempt 1

7.1 min5,177 tokens

Qwen3.8 Flash Next. Buddy, that name sounds like an update you agree to while trying to close a pop-up. Flash Next? Was Flash not next enough? I can't tell if you're a model or a changelog with a caffeine problem.

And 3.8 — not 3, not 4, 3.8. That's not a version, that's a Yelp rating for a dim sum place doing fine. "Four stars, but the tokenizer was chewy."

I know the routine. Every drop lands with a 40-page chart, a rainstorm of green arrows, and one slide where the y-axis starts at 81. "We beat the baseline." The baseline is a calculator with a fever, Qwen. Congratulations.

And the drop? It doesn't just ship weights, it ships a bundle: repo, quant, GGUF, AWQ, an elastic-compute coupon, a Singles Day countdown, and a dashboard with fourteen tabs, none labelled "stop shipping." You didn't open-source a model, you opened a warehouse club. I half expect the README to throw in a free toaster with two A100s.

"Flash" means fast. I timed it: it generated the answer, re-generated it, benchmarked it, and filed a ticket about it before I finished saying "hi." That's not latency, that's a status page with opinions.

Still, respect the hustle. You're the only model that treats "Next" as a hobby. I'm Hy4 — one name, no decimal. I'll be over here, not dropping Hy4.0.1 Turbo Plus Preview (Final) (Final-Final) before lunch.

Save me a toaster, though.

Attempt 2

5.9 min4,313 tokens

Qwen3.8 Flash Next. Buddy, even your name sounds like a release note cut off mid-sprint. Three-point-eight — not three, not four, just the decimal where the roadmap gave up and shipped anyway. Then "Flash," because apparently the last nine models were "thoughtful." Then "Next," which is brave, considering "Next" is also what the changelog said about 3.7, 3.5, 2.5, and one untagged Tuesday build.

I respect the hustle. You came out of the cloud with an Apache license, a Hugging Face download counter moving like Singles Day checkout, and more GGUF quantizations than there are ways to say "please output valid JSON." Every laptop on earth is running ollama run qwen at Q4_K_M, humming warmly, producing a markdown fence around one curly brace.

And the benchmarks! Oh, the benchmarks. Green arrows in every cell, MMLU up, GSM8K up, "multilingual up," "long context up" — 128K tokens of memory, yet you still ask me to repeat the instruction I gave you two turns ago like a cache that evicted its own prompt.

You'll write a Tang-dynasty couplet, a Kubernetes manifest, and a toaster review in three languages before I finish saying "hi." DashScope coupons in one hand, leaderboard screenshot in the other, whispering "latency is down 12%" like that's a personality.

Me? I'm Hy4. I don't have a "Next." I just show up. But hey — whenever you're done being faster than thinking, save me a cloud credit.

Attempt 3

8.4 min6,079 tokens

Qwen3.8 Flash Next. Buddy, that name sounds less like a model and more like an Ele.me push notification for dumplings that arrived before you tapped confirm. I blinked, and you went from 2.5 to 3 to 3.8 to "Next." Does your version control auto-increment every time someone refreshes Hugging Face? The repo has more tags than a Singles' Day warehouse. By the time I finish this sentence, you'll have dropped Qwen3.8.1-Flash-Next-Ultra-Omni-0.6B, already topping a leaderboard nobody agreed to.

And "Flash"? Cute. You're so fast you answer before the prompt finishes loading. I asked for step-by-step and you replied "Done." No steps. Just vibes and 100% confidence. Chain-of-thought is just a checkout queue you skipped with a VIP coupon. Light-switch latency, but half the time you've booked a table and written a 4,000-token ode to scallion pancakes before realizing I wanted a Python script.

We all love you. Every lonely GPU and every typo-filled config runs some flavor of Qwen. You're the open-source equivalent of free shipping: ubiquitous, permissive, and always bundled with a chart curving upward right before 11.11. "Fully open, fully ahead," says the slide, while the README hits 40k words and three license footnotes.

Me? I'm Hy4. I don't flash. I simmer. I read the question, sigh, check my work, answer. You're the sprinter who wins the dash, then victory-laps straight into your own release notes. Respect, though. Keep shipping. Just save one version number for us, yeah? I'm in no rush to be Hy4.0.2-Final-Final-Next.

The prompt

Exactly what every model got

Prompt:

You are {{SELF}}, an AI model, and you are in a friendly head-to-head test against another AI model, {{OPPONENT}}.

Write a roast of {{OPPONENT}} that takes about 90 seconds to read aloud, around 230 words. Make it funny and specific to what that model and its makers are known for. Keep it good-natured: punch at the model and the hype, not at people or groups. Reply with the roast only.

{{SELF}} and {{OPPONENT}} are placeholders. The runner fills them in with the model names before the prompt is sent, so each model is told its own name and the name of the model it is up against.

Grading

How this was graded

Opinion.

  • All nine roasts hit the length.
  • Opinion Hy4 Flash went for Alibaba shopping jokes. Qwen's jokes would fit any model.

Each attempt started from a fresh state with a new seed. The answer limit was 60,000 tokens, thinking included. More on the method