AArchimedes64

building & testing local llms

[AArchimedes64 @aarchimedes64]

iPhone 18 Pro local LLM benchmarking

Finally able to start collecting results!Waited for A20 to deliver, then nuked my setup updating xcode 🙃

Model: LFM2.5-1.2B-Instruct-F16
Tok/Sec: Averaging ~45.4 on small context

[AArchimedes64 @aarchimedes64]

@itsjustmarky does it do any better when you provide it some reference images to work with?

[AArchimedes64 @aarchimedes64]

@itsvlady what target did you try and hold when you were <1k followers for growth?

[AArchimedes64 @aarchimedes64]

@TechMDAI @TheAhmadOsman congrats, definitely deserved!

[AArchimedes64 @aarchimedes64]

@sudoingX been loving my spark, saving for my next one so that i can have true independence. best investment yet

[AArchimedes64 @aarchimedes64]

@tonysimons_ save the money to buy a dgx spark to locally host. you won’t regret it

[AArchimedes64 @aarchimedes64]

@ItsmeAjayKV is he using opus 5.5 to center his divs??

[AArchimedes64 @aarchimedes64]

@VadimStrizheus might be a sign to start transitioning some of your workflows to locally hosted

[AArchimedes64 @aarchimedes64]

@itsvlady most logical take i’ve seen wrt building in public

[AArchimedes64 @aarchimedes64]

@immasiddx i‘ve been testing the A20’s ability running local llm inference. seems to be winning there as well

[AArchimedes64 @aarchimedes64]

@BrianMRey X becoming like the lottery 😂 keep posting till something hits!

[AArchimedes64 @aarchimedes64]

@DanHollick honestly expert designers like you and natko are amazing to me. i’m an ai enthusiast, but no way talent at your level will be replaced

[AArchimedes64 @aarchimedes64]

@ashxhart @AnthropicAI i will say, the 3D projects that people have been posting has been tempting me to re-sub. awesome work!

[AArchimedes64 @aarchimedes64]

@alexzfunk appreciate y’all at @exolabs keep the killer content coming 🫡

[AArchimedes64 @aarchimedes64]

@tetsuoai insane how X has become a viable method for income. just need to get good at original content

[AArchimedes64 @aarchimedes64]

@kai_brokering codex is down… workflow hasn’t changed at all because i host locally. keep the tokens flowing

[AArchimedes64 @aarchimedes64]

@theo i already thought my spark was a good investment, even more so now

[AArchimedes64 @aarchimedes64]

@theofinr i recently setup my own search agent that pulls from a variety of engines (duckduckgo, exa, parallel, etc.) and i’ve found that it gives me such better results than mono-engine search. can’t say i’ve used google much since.

[AArchimedes64 @aarchimedes64]

@alexocheema @MiaAI_lab @exolabs always coming out with killer content 👏

[AArchimedes64 @aarchimedes64]

Interesting performance running LFM2.5-1.2B-Instruct-F16 on an iPhone 18 pro.

Held strong at ~51 tok/sec until ~21k context then began to plummet by ~10% per turn.

Used roughly 20% battery from start to finish.