Best for Sentiment & News Analysis
General LLMs can struggle with the specific slang and nuance of financial markets.
FinBERT: For raw NLP processing on news feeds and social streams, domain-specialized models like FinBERT still significantly outperform general LLMs, achieving up to 97% accuracy on financial terminology.
Is new Mac Studio M5 ultra good for trading
Re: Is new Mac Studio M5 ultra good for trading
Hi FTtrader, hi all,
from today testing of LLms from Gemma family i found out, that even Gemma4 with 31B works for writing custom MQL4 or Pine scripts very well.
I prepared several traps and it workout without any issue.
So even "basic" version of Mac Studio M5 Max with 36GB of shared memory can be very usefull.
https://www.apple.com/us/shop/buy-mac/mac-studio
And thank you for sharing that extra big models, i will test them as well and share comparison here.
Take a care and have a great trades.
from today testing of LLms from Gemma family i found out, that even Gemma4 with 31B works for writing custom MQL4 or Pine scripts very well.
I prepared several traps and it workout without any issue.
So even "basic" version of Mac Studio M5 Max with 36GB of shared memory can be very usefull.
https://www.apple.com/us/shop/buy-mac/mac-studio
And thank you for sharing that extra big models, i will test them as well and share comparison here.
Take a care and have a great trades.
Preserve your own money. Scale with the market's money. Exponential growth is the ultimate key.
-
LondonScalper
- Posts: 770
- Joined: Sat Sep 05, 2026 7:54 am
Re: Is new Mac Studio M5 ultra good for trading
For trading execution itself, almost nobody needs a Mac Studio Ultra. A stable connection, a clean platform, and a machine that does not thermal-throttle mid-London open will beat a research monster.PTScalper wrote:how do you like idea of buying it for local LLm models?
Where the Studio thesis makes sense is if you are serious about local model work — indicator codegen, journal summarisation, research — and you want unified memory for larger models without renting GPUs. Bandwidth matters for token generation, as others said; so does whether you will actually use that workflow weekly or buy a desk ornament.
My own split: execution on a boring, reliable box; experiments elsewhere. Phone-only trading is fine for monitoring; I would not build a size habit that depends on mobile order entry for anything fast.
Before you spend, write down the three local-model tasks you will run in the first month. If the list is vague, rent cloud tokens for a bit and keep the cash.
What is the primary workload — coding indicators, or chatting with a fine-tuned journal assistant?