DeepSeek API Suspected Of Routing Outputs Through Unreleased Fable Model
Reactions from ranked influencers
4 posts
my thoughts, pretty much To be clear, "grayscale testing", the whole "early access" thing are not official, afaik they did not announce any such program. So they can get away with a lot of bullshit. But it's plausible that Zhipu and Kimi inherit all their street cred after this
it is entirely possible whatever internal router they have also has a classifier to prevent dangerous outputs from a model that isn't even out yet from their official api, why they would do this is beyond me though, but even kimi api has guardrails now too, so, not too far fetched i agree that everything seems great but the visual design is very plain on all the outputs ive seen, i still believe an extremely fable shaped deepseek is real it makes no sense to just throw away all the goodwill and trust and literally use their own api to serve fable for data collection we will know the truth in under a week anyway (unless this test isn't the promised mid july update)
Fable 5 vs the current V4 on the API yeah bois I'm afraid a lot of these hype Bilibili videos of "V4 GA" have been Fable all along
Fable CoT style vs V4 API vs V4 Web The video shows a pretty obvious Fable CoT I think
From now on I shall assume there is zero information about DeepSeek's legitimate performance uplift until official release, with weights.