6,756 karma · joined October 14, 2022
Do you perhaps have a link to it? Tried to search around but couldn’t find it. Would love to read it
You can always try out this model for free on Qwen Chat. Alibaba Cloud has too much friction for me.
Update: It clicked now…
See Kimi Vendor Verifier.
Kimi had become that popular. I was a subscriber of Kimi back when latest version was Kimi K2. Later I unsubscribed because I jumped over to GLM subscription (they had amazing deal). Now when I wanted to try out Kimi K3 to find out what the fuzz was all about, I couldn’t subscribe to them.
I remember reading a post from Moonshot team about this, they are doing this because they are almost at peak capacity and want to reserve it to keep the quality for their current customers.
We are actually witnessing an open-weight model catching up at catching mainstream users attention. And instead of behaving like Anthropic, they actually care about their users experience.
This is what the Chinese always been good at. Take expensive innovation and streamline it to lower prices. But we are at a point where labs like Moonshot actually contributes a lot to the research field as well. They are pushing the innovation forward and squeezing the prices. Very well done.
Whats even weirder is the bizarre mechanisms Anthropic implemented to prevent distills which they had to sacrifice their customers for. They hid the internal CoT reasoning and returns summarizations instead. This made it difficult for users to trace things. They made Fable 5 silently switched over to Opus 4.8 if it detected blacklisted prompts (almost anything triggered this) to sabotage distills. And now, they are still complaining about distills? So their customers have gotten sacrificed over nothing.
Whats even weirder is the timeframe here, no way the Moonshot team managed to plan conduct a large scale distill, then pre-train, RL, fine-tune, benchmark, marketing and release to their platform since Fable 5 got whitelisted.
> they developed a sophisticated internal platform to conduct large scale distillation
I am very curious about this and would love to learn more on how they did this. Wish we had more details. I know the team behind DeepSeek have also done clever things to distill too. I am aware of these ”transfer stations” that acts as a proxy, but I don’t think they are helpful in this case.
How will this affect OpenAI?
Otherwise, this news feels like a tiny incremental improvement on Gemini Flash series to make it more efficient with token usage, subagent and cost. Nothing big.
Regarding their benchmark scores on CyberGym, I wonder why they didn't compare their 3.5 Flash Cyber model with Fable 5. I mean they included Mythos and GPT-Cyber, so why not Fable 5 too?
They also mentioned Gemini 3.5 Pro is in testing and its about to become available very soon. Another thing maybe worth discussing is the announcement of pre-training Gemini 4. Sadly, not much technical details to discuss on. Many comments in here seem to mostly be about how Google is behind the others, but honestly, is it really worth the investment to be #1 in Artifical Analysis every week?
Why do you care that much? Just say it. You're letting an imaginary score determine if you should express your suggestion for improvement. That's a bit wild.