Kimi has sound, it's just very quiet
The original k3 also matched its respective opus
This seems more a measure of anti promp injection than general intelligence
I hope someone runs multilangbench for other models. I really want to see how open models in particular do on it. Benchmaxxing by focusing on the top 5 languages seems an easy
That's fascinating. Are these for your keyboard? You may have gone too fast honestly. Something like parakeet runs fast enough and seems to have much lower wer
Which claude model are you comparing it to? If you're using fable max obv it's going to be slower, but like sonnet high is both faster and more intelligent
Very similar to Kimi K3 it seems, but fewer active params (80b vs 104b). I don't think they mentioned what form of attention they use
This is why I severely dislike using all pass rate. A model could be at like 93%, meeting practically every requirement except the broken ones and all pass would give it like 20%.
I use it for title generation basically. Will have to see where this one fits in
Try fable. I haven't used it since they dropped it from the pro plan, but when I did, fable 5 casually dropped such advanced electrical and orbital mechanics knowledge in my story that I had to stop and ask it to explain
The biggest discovery might actually be that they ignored regular caching rules and kept invalid cache suffixes and it didn't hurt performance
Look at Neuralwatt. They report energy usage with every call as well as aggregate statistics.
Why are you assuming the gpus are rented? Anyway, they still have token based pricing if you prefer. The energy pricing is often cheaper though
This article feels like it was written by a human who has just read to much llm writing
I think it's more likely they recorded the video when the project was already done than cheated
Why?
Alibaba is entirely capitalistic I'm afraid
People often use this a argument against small phones like this wouldn't be amazing for any company outside apple or Samsung
What if you don't own a phone?
The intermediate tickers are fake but real data comes in and resets it. Its like a progress bar essentially. We don't call progress and bars fake
I don't really remember a situation, which of those models supposedly beat the other? I still opus 4.6 though not for code
Distillation requires you to have the actual logits of each token from the teacher model, which in practice means having the model itself. What you're describing is just synthetic data. Note Anthropic misused the term…
This and their newer Aura A1 I've wanted to get cause of the compact size, but they seem to lack support for US 5g bands
I highly doubt it's behind in practice, except for Anthropic
Cache hit % on openrouter is not a good metric, it's mainly driven by openrouter's own provider juggling than the providers themselves
Kimi has sound, it's just very quiet
The original k3 also matched its respective opus
This seems more a measure of anti promp injection than general intelligence
I hope someone runs multilangbench for other models. I really want to see how open models in particular do on it. Benchmaxxing by focusing on the top 5 languages seems an easy
That's fascinating. Are these for your keyboard? You may have gone too fast honestly. Something like parakeet runs fast enough and seems to have much lower wer
Which claude model are you comparing it to? If you're using fable max obv it's going to be slower, but like sonnet high is both faster and more intelligent
Very similar to Kimi K3 it seems, but fewer active params (80b vs 104b). I don't think they mentioned what form of attention they use
This is why I severely dislike using all pass rate. A model could be at like 93%, meeting practically every requirement except the broken ones and all pass would give it like 20%.
I use it for title generation basically. Will have to see where this one fits in
Try fable. I haven't used it since they dropped it from the pro plan, but when I did, fable 5 casually dropped such advanced electrical and orbital mechanics knowledge in my story that I had to stop and ask it to explain
The biggest discovery might actually be that they ignored regular caching rules and kept invalid cache suffixes and it didn't hurt performance
Look at Neuralwatt. They report energy usage with every call as well as aggregate statistics.
Why are you assuming the gpus are rented? Anyway, they still have token based pricing if you prefer. The energy pricing is often cheaper though
This article feels like it was written by a human who has just read to much llm writing
I think it's more likely they recorded the video when the project was already done than cheated
Why?
Alibaba is entirely capitalistic I'm afraid
People often use this a argument against small phones like this wouldn't be amazing for any company outside apple or Samsung
What if you don't own a phone?
The intermediate tickers are fake but real data comes in and resets it. Its like a progress bar essentially. We don't call progress and bars fake
I don't really remember a situation, which of those models supposedly beat the other? I still opus 4.6 though not for code
Distillation requires you to have the actual logits of each token from the teacher model, which in practice means having the model itself. What you're describing is just synthetic data. Note Anthropic misused the term…
This and their newer Aura A1 I've wanted to get cause of the compact size, but they seem to lack support for US 5g bands
I highly doubt it's behind in practice, except for Anthropic
Cache hit % on openrouter is not a good metric, it's mainly driven by openrouter's own provider juggling than the providers themselves