Video yükleniyor...
Video Yüklenemedi
Generally, free AI comes with weird limits. However, with this stealth AI model, Ox Alpha + OpenCode Zen gateway, I can personally verify that it is: ✅ Almost unlimited running our internal cybersec bench (I burned 4 billion tokens in 4 hours) ✅ Decent speed ✅ Intelligent model (solves... show more
39,953 görüntüleme • 26 gün önce •via X (Twitter)
38 Yorum
It's mostly likely either of these: -GLM model -Hy4 -MiMo V3 -MiniMax
I’ll be surprised if this is not an American lab model
@1kartikkabadi1 how would an american frontier lab be able to run cybersec with no guardrails, look what happened with fable. why would gemini be above them?
@1kartikkabadi1 Grok 4.6 also has no guardrails as such. No guardrails don’t mean that model is SOTA in cyber security tasks (5.6 sol crushes this model in our bench)
It's a chinese model
Running your internal cybersec bench on a free stealth model is ballsy. Hope that zero retention claim comes with audit rights.
Like our tasks we sell labs, this bench includes offensive multi service web security RL environment tasks. AI can’t see/exfil environment code
Ok fair. Exfil angle is covered. Still wonder where those 4B tokens end up.
Leaning more towards it being a Chinese model than Gemini 3.5 pro, Testing it out right now
gemini doesn't have max thinking level, probably minimax new model, if not, xiaomi
Four billion tokens in four hours confirms these are real numbers. We actually went deeper on what makes it genuinely different here:
4B tokens in 4 hours with zero guardrails sounds less like a standard gateway leak and more like an unthrottled red-teaming sandbox in the wild. The throughput is genuinely insane though.
It’s new GLM model.
Would be funny if it is Gemma 5
Oh shit. Lol. Nobody even bothers to think that this could be a google model. And you are right. Who has this much capacity to serve their model? Ofcourse that's google. I think u are the only one I have seen on my timeline mention this being a google model. Let's see next week.
Comparable to deepseek v4 flash in outputs?
É chinês amigo!
Deepseek V5 Flash for sure
It's already proven the model has the exact same tokenizer and video encoder footprint as GLM models... why would American labs has the same tokenizer as GLM?
Pretty humble model NGL
It's GLM
Idk man, they have so much compute 100 trillion is a fucking huge number
Were you always this dumb it's glm 5.3
Gonna Try 0x Alpha for my project from GPT Luna
Not Gemini. I think I learned while working with Gemini models is that they have the best support for regional languages like Malayalam that other models don’t. Prompt it with obscure Malayalam words, and you will know the difference.
what's the actual per-request token limit or context window you hit
It’s astra btw
My guess -
lmao🤣
Why would Gemini 3.5 Pro not have Cybersecurity Guardrails?
It's so crazyyy I'm gonna try it today
No one else thinks it’s Google
Same here burned through 1B tokens and worked as well/better than sol for most the tasks
I mean google did launch nano-banana like this, although the difference is, everybody was impressed with it, unlike this one.
Can I use in claude?
But we can see the thinking does gemini models shows thinking texts?
Interesting
Its GLM 5.3 Vision
