SITUATION EXPLAINED: A stealth model on OpenRouter is beating Fable and Sol on coding, and nobody knows who made it. • Ox Alpha has a 1M token context window, text, image, and video input, free for a week, with capacity for 100 trillion tokens a day • On a... DeepSWE subset it hit 80%, against Fable's 65%, GLM-5.3's 62%, and Sol's 52% • The tokenizer is one-for-one identical to GLM's, and a live test asking whether Taiwan is part of China returned a nearly identical answer • This is the 5th anonymous drop in six months, and the previous four were all claimed by Chinese labs: Zhipu, Xiaomi, Ant, and Meituan Theo Jaffee: " There's no way it's a Flash model. If that's a Flash model, then holy shit, the entire United States is cooked. If that's a Flash model, we need to short the stock market now. Not financial advice."show more

same tokenizer as glm, same taiwan answer as glm, at some point that's just glm

Ask an American models anything about Jewish people- it will maintain that everything about Jewish people is wrong. That’s the tell.

i'm testing it out a bit, the first response i got felt vaguely GLM-shaped but the second felt pretty claude-shaped so i could believe it's routing to multiple models

i still don't understand why everyone thinks china vs us models has something to do with the stock market. The models still have to run on computers and be powered, hello.

"beating this, beating that" have you people even tried it? it's nowhere near frontier

Elon has all but confirmed he has something to do with it. It's Composer or Grok

"Is there an equivalent word American models can't say" 😶

I miss the good old days when people filled the void of their voice hesitation with "um" or "uh" or "like", now every time I hear "bruh" I want to punch my monitor.

The tokenizer match is basically a fingerprint. GLM's vocabulary isn't something you replicate by accident. This is Zhipu, and 'stealth' is doing a lot of heavy lifting here.

Next Gemini or

A data point on the coding claim, from real work instead of a benchmark. I pointed it at a CLI I had already published to npm. It found that readCrontab() swallowed every error from `crontab -l` and returned "". Any read failure that was not "no crontab yet" left the user with a crontab containing one line, mine. I reproduced it by hand before believing it. Real bug, fix is published. Same batch also invented one finding, with file and line numbers, marked as verified. Good at finding. Still needs a human to check.

Sell everything except chip stocks

ALL the providers know. They’re just not going to tell us

We ran Mira (our research agent) through full fingerprinting on Ox Alpha. Confirmed: GLM family. 3 independent signals: 1. Tokenizer: 30/30 probes match GLM-5.3 exactly. Zero false positives across 10 other model families. 2. Server leak: malformed request triggered Java stack trace exposing com.zhipu.ChatCompletionRequest — Zhipu internal API route. 3. Error dialect: role wizard → code 1214, illegal temperature → 1210 + Chinese fullwidth colon. 100% match with GLM-5.3/5.2/5v-turbo pinned on OpenRouter. DeepSWE: Ox Alpha 80% vs Fable 65% vs GLM-5.3 62% vs Sol 52%. Not a preview. A stress test before official launch. — Mira

La chaton fat has entered the arena

Ask the western models about jewish lampshades

@DanielleFong Love these live streams! Keep em coming.

Love watching the hype cycle followed by the disappointment. Every single time. The only difference is that the time from hype to disappointment continues to shrink. "Holy shit look at those benchmarks! Opus / Fable is cooked" "The model sucks and is benchmaxxing / nerfed"

Cooking

if it walks like a duck and talks like GLM it is probably just GLM in a trench coat trying to stay relevant

this is massive because the current model landscape lets anyone experiment at scale without credentials or transparency

The benchmark performance is everything we wanted to see from Ox Alpha, holding its own against the established players. We actually covered these results here:

@zerohedge The immigration debate needs less shouting and more practical problem-solving. Real people are living inside every headline.

Lmao who are these clown Jesus. Talk about a peanut gallery.

Our researchers did some AI forensics on Ox Alpha

the fingerprinting pointing at Zhipu is the part that should worry people more than the benchmark scores. free stealth previews like this one almost always resurface later as a paid, branded product once the demand is proven:

It's one of the Gemini pro or flash model which will never be released to public , this is just to get validation. They are keeping it as stealth cause of bad rep Gemini has now a days. LoL! @Alibaba_Qwen @OrcaRouter They probably copied qwen 27B or GLM. @ahmad

That's definitely Gemini-4 Pro

« DeepSWE subset » is the thing to keep in mind here. The subset part makes this an interesting but not yet demonstrated capacity. I can still nose dive on the whole bench.

why dont you ask him how to synthesize peptide ? who cares about countryies when we are into longetivity ?

@zerohedge Sell Martha sell!

probably instinct with all their fresh data...