Just tried it on a medium size coding/debug problem on an existing codebase, observations: - Input doesn't look faster than other models, it spends a lot of time reading Read about 5M tokens - Output is awesome, super fast as you expect from the 1500t/sec I think that's correct - Tool call is failing more than say DS4, which leads to time wasted on retries (complex tools like browser control for example) - Shell commands are still somewhat of a bottleneck
The net effect is that I spend about the same time waiting, and I still need to read that output so, at least for coding, it actually reconciles me with the 100-200t/sec you can get on DS4 or the like. Maybe that's a good sweet spot after all and faster t/sec is not where the bottleneck is.
Also maybe my setup (OMP) doesn't do the cache correctly but that's a huge cost driver... so atm it's quite pricy
> "Also maybe my setup (OMP) doesn't do the cache correctly but that's a huge cost driver... so atm it's quite pricy"
I don't believe Cerebras has a cached input pricing? They don't list one on the model page:
https://inference-docs.cerebras.ai/models/qwen-3.8-27b
edit: See the sibling discussion,
https://news.ycombinator.com/item?id=49554520#49555094 ("Input tokens, whether served from the cache or processed fresh, are billed at the standard input token rate")
lol yeah just saw that, yeah that makes it unusable I think at least for me.
I wonder if they will do that with sol ultrafast!
They have cache, but it costs the same indeed, no idea what the point of the cache is
They don't have cache (e.g. KV cache). But they write down what you sent earlier to say they cached it! To still bill the same as uncached later (because they didn't actually cache it)!
More precisely they can't cache it.
Yes. Their architecture recomputes every time so at 150k context every request will have to spend 1.5 min waiting for the model to reread the context.
Say avg model response length is 1024 tok. At 50 tok/s normal providers do your turn will only take 20s (vs Cerebras 101s) and will cost 20x less. That time and cost is per single tool call.
I can't believe this situation has not improved in years. Is cerebras' main business selling the hardware, then?
they have exactly two customers, both of whom are also investors.
Maybe they’re gunning for speedy non-interactive pricing? Or its a limit of the technology or a business decision?
Cerebras is super constrained on capacity right now, all the support is going to enterprise customers.
[dead]