Gemini 2.5 pro / Deep Think VS local LLM
Posted by Dumperandumper@reddit | LocalLLaMA | View on Reddit | 25 comments
I’m on « Ultra » plan with google since 3 months now and while I was cool with their discovery offer (149€/ month) I have now 3 days left to cancel before they start charging me 279€/ month. I did heavily use 2.5 pro and Deep Think for creative writing, brainstorming critical law related questions. I do not code. I have to admit Gemini has been a huge gain in productivity but 279€/ month is such a heavy price just to have access to Deep Think. My question is : are there any local LLM that I can run, even slowly, on my hardware that are good enough compared to what I have been used to ? I’ve got a macbook pro M3 max 128gb ram. How well can I do ? Any pointer greatly appreciated. Apologies for my english. Frenchman here
AlgorithmicMuse@reddit
Nothing you run locally will compare to the cloud based llms
quanhua92@reddit
Why don't you downgrade to cheaper plan? I use Gemini 2.5 Pro with $20 plan. I think Ultra is only useful if you want to use lots of image and video generation.
You can try to use the Google AI Studio to run Gemini 2.5 Pro for free as well.
For Local LLM, you can try LM Studio and download some common big models like gpt oss, qwen3, gml 4.6. However, I think you will need the cloud plan for Deep Research anyway. Using local LLM with web search API is not cheap.
So, my suggestion is to use cheaper plan first. Then switch to Local LLM when you hit the rate limit.
Dumperandumper@reddit (OP)
Good question. I do need to put out alot of creative writing for a living, and Deep Think litterally kills 2.5 pro in that field, despite being limited to 10 queries a day. I also study real world law cases for some of my writing and again 2.5 pro is lacking big time behind Deep Think. I’ll check local LLM options or go with OpenAi or elsewhere ! Thanks for your feedback
rz2000@reddit
Consider Kagi Assistant with its full menu of AIs including Gemini, Claude, ChatGPT, DeepSeek, Kimi, GLM etc as well as integrated Kagi search for legal research.
It might be a way to try out the different models to see what you want to try to run locally.
I use local models on a powerful workstation and supplement them with cheaper $20 Gemini and Claude plans as well as Kagi Assistant. Especially when researching legal topics it can be useful to let a few different AIs run with deep dives, then compare them to get a better idea as to whether any of their thought processes are actually kind of insane.
quanhua92@reddit
I think OpenAI will cost you $200 as well. May be Claude Max $100?
TumbleweedDeep825@reddit
sign up for a trial under a new gmail account. then you're back at 149.
Steus_au@reddit
openrouter - you could have about 50 million tokens per months for this money
PathIntelligent7082@reddit
openrouter is crap...they cheat
AlbanySteamedHams@reddit
could you elaborate?
Fall-IDE-Admin@reddit
O don't think running a local llm will work. I instead suggest to use one of the open source deep research projects on GitHub and use them with a llm service of your choice.
taoyx@reddit
Download LM Studio and use your 3 days to compare Gemini pro vs some LLMs like Qwen or Mistral. I don't know what are the best for your configuration LM Studio should sort that for you.
dhamaniasad@reddit
ChatGPT has a deep think parallel with their pro models. And it’s cheaper than Gemini by a bit. No local model you can run, nor any open source model, will come close to the performance of these models. You have some models that can match the frontier models, like 2.5 Pro, I think GLM 4.5/4.6, the larger Qwen models, etc. And you can use OptiLLM to get something similar to the Pro / Deep Think modes of these models. GPT-5 Pro is available on the API if your usage is not enough to justify $200 per month but know that it racks up costs very quickly on the API.
ParthProLegend@reddit
Any Local LLM you run on Macbook or AMD platforms are NOT close to the likes of Gemini 2.5Pro/Deep Think. So if you expect a 20% loss at max in quality, go with Qwen 3 80B
Dumperandumper@reddit (OP)
I kinda knew M series GPU are weak vs Nvidia. 20% loss, I guess I can try ! Thanks for your reply
Eden1506@reddit
qwen moe models are terrible at creative writing btw
try glm 4.5 air instead it has a finetune for creative writing by drummer called GLM Steam
ParthProLegend@reddit
They are decent.
Eden1506@reddit
qwen moe models are terrible at creative writing btw
Eden1506@reddit
Something like glm 4.5 air 106b with websearch should run and give you around 2/3 of what you are after.
asankhs@reddit
You can try using MARS plugin in OptiLLM - https://www.reddit.com/r/optillm/comments/1nwx307/mars_in_optillm_73_on_aime_2025_with_multiagent/
Character_Act7116@reddit
qwen3-next-80b MLX
Steus_au@reddit
it would be painfully slow on large prompts though
power97992@reddit
I dont think there is anything comparable to deep think even if you had 1tb of vram …. You are better off switching to gpt 5 pro ….
chisleu@reddit
Bonjour.
Short answer, you will get really really close but not quite there will local LLMs. You have a fantastic platform. Download LMStudio and download the largest, most recent recommended model that your system can run. There are a lot of options to choose from. That's likely going to be GPT OSS 120b.
Dumperandumper@reddit (OP)
Bonjour and thanks. Exactly what I needed to know. I’m gonna try this and see how it runs
SM8085@reddit
gpt-oss-120b-GGUF