Could Ox Alpha actually be Le Chaton Fat/Le gros chaton?https://openrouter.ai/stealth/ox-alpha
>>109615431>30T parameteri'm sorry i'm not an AI troon trump voter, but isn't 30T like a couple of those nvidia server class gpu banks?
>>109615431anybody's guess; people are saying Ox Alpha won't talk about tiananmensquare, others say it does but refuses cybersec
nah it's some chinkshit because it won't talk about the usual chink censored subjects (tianamen, xinjiang, etc)>>109615465it's a meme
>>109615431Europe doesn't have the infrastructure for that, it's not possible. Also probably illegal to stealth deploy a model.
It's GLM 5.3 Air or Qwen 3.8 122B.
Mistral will release another 120B model in December which will rival the ranks of Qwen3.5-9B
>>109615504>people are saying Ox Alpha won't talk about tiananmensquareit got this far before shitting the bed
>This stealth model is developed and operated by a third-party model provider. Prompts and completions for this model are retained by the provider and are not used for training; all other use is governed by the Stealth Model Terms(opens in new tab).what does an AI lab gain by offering their model like this for free for a week? are the incoming prompts that valuable? what's the plan, replaying these prompts to claude for a more broad distillation?
>>109615690they manage to get in-distribution traces, e.g. what most normies use their coding agents to automate and anything flagged by a super cheap classifier as failed, gets insta pushed onto the RL-env-benchmaxxing setup to create envs to hill-climb on it.t. does this daily
>>109615504>>109615506>>109615603Works on my machine.
>>109615431That cat looks like buffcat but fat
>>109615777>anything flagged by a super cheap classifier as failed, gets insta pushed onto the RL-env-benchmaxxing setup to create envs to hill-climb on ithow does training work in this case if there's no ground truth correct answer available? is the classifier really so good it can tell when the model manages to solve the problem correctly?
It's a GLM model. It has the exact same writing quirks and my GLM-focused jailbreak prefill works with it.
>>109617181But wouldn't including Tiananmen Square and the treatment of Uyghurs (see >>109616186) in the training data indicate that it's a Western model trained on a GLM base? Mistral has started to serve GLM 5.2: https://mistral.ai/news/regional-inference-open-models-new-compute/The other candidates are SpaceXAI and Cursor.
>>109615504>chaton>>109615506>>109615522>>109615603>>109617181Looks like you are going to be right.https://nitter.net/MiaAI_lab/status/2091150669323346144
>>109619979>Has this been verified?>Yes>(Source: random xitter schizo)
Mistral doesn't have the resources to compete with American AI and they aren't going to be as cheap as the Chinese ones. Starting any type of tech business in the EU or a place like France with endless data regulations and bureaucracy is already a handicap
All the evidence supports that ox alpha is glm-5.3-flash being served by one or more of the US neocloudsIf they release the weights (which they probably will since they already distributed the model to the providers) you will likely be able to run this thing on 2x dgx sparks
>>109615431>Could [LITERAL WHO MODEL] BE [LE FRENCH MEME CAT]???????This is basically an AI Twitter post. Fuck off.
>>109620352The American inference provider is most likely Fireworks, right?
