South Korea says it's uncovered evidence that DeepSeek has secretly been sharing data with ByteDance, the parent company of popular social media app TikTok.
This is the web chat client/app, just like OpenAI sharing data with Microsoft, or Copilot doing the same. If you self host these LLMs your data stays within your LAN.
I’m moving to self host all my streaming stuff. Switching from local-only plex to self hosting all my media (spotify, google photos, LLMs) and tools behind a reverse proxy so i can access outside my home. It’s pretty sweet and a good learning experience using reverse proxies
Edit: Plus fuck these technofeudal lords who enclose access to markets, information, and culture.
Do you know of a provider is actually private? The few privacy policies I checked all had something like “We might keep some of your data for some time for anti-abuse or other reasons”…
I dont really use LLMs so I didn’t even realize there were versions with different weights and stuff. I was using 7b, but found it pretty useless. Pretty sure I’m not going to be able run 32B on my rig. lmao.
I’m not @[email protected] However here’s a pretty barebones how to article to get you started. Just know it can be as complicated as you like. For starters you may want to stick to the 7b and 14b models like mistral:7b and phi4:14b as they’ll fit easily on your card and will allow you to test the waters.
Locally? Arcee 14B and the 14B Deepseek distill are currently the best models that fill fit.
I’d recommend hosting them with TabbyAPI instead of ollama, as they will be much faster and more VRAM efficient. But this is more fuss.
Honestly, I would just try free APIs like Gemini, Groq, and such through open web ui, or use really cheap APIs like openrouter. Newer 14B models are okay, but they’re definitely lacking that “encyclopedic intelligence” larger models have.
I use 32b and the 672b side by side. The performance hit is around 20% and I keep all my data local. I am not conflating the two however self hosting works for me just fine. Your usecase is your own certainly. However I’d rather take the performance hit for the added data privacy.
Also it’s nice to he able to set my own weights and further distil R1
I have a local python expert a local golang expert and both have my local gitlab repository and I’ve tied their respective Ollama keys to my VSCode IDE.
With the distilled models I have, I’ve been able to build and troubleshoot pretty complicated apps in Golang and Python. However, these distilled models are very specialized and will not do things like write me a story about a duck made out of duct tape or properly summarize articles. There are absolutely limits to my workflow and setup. But I’m pretty happy with it.
Have you had any luck importing even a medium-sized codebase and doing reasoning on it? All of my experiments start to show subtle errors past 2k tokens, and at 5k tokens the errors become significant. Any attempt to ingest and process a decent-sized project (say 20k SLOC plus tooling/overhead/config) has been useless, even on models that “should” have a good-enough sized context window.
My codebase is almost 1.2GB of raw python and go files no images. I think it’s somewhere near 15k tokens for the python codebase and 22k for golang due to all the .mod and .io connectors to python libraries… it was a much bigger mess before if you can believe it.
What size model are you using? I’m getting pretty good results with R1 32b but these have been distilled to be experts in the languages of the codebases. I’m not using any general models for this.
Also it depends on the language you’re targeting as well. Rust or Lisp have issues due to how much less they’ve been documented. I think golf type languages like brainfuck are impossible. It really comes down to how the language has been documented. Python gave me issues in the beginning until I specified 3.11 in my weights and distillation/training, and that definitely fixed a lot of the hallucinations I was getting from the model.
I think static typing languages that have consistent documentation would be the easiest for this. Now that I think of it, maybe getting a typescript expert would be something I could tool around with.
Edited for legibility and the fact that I just went and looked at my datasets again. Much bigger than I initially thought.
This is the web chat client/app, just like OpenAI sharing data with Microsoft, or Copilot doing the same. If you self host these LLMs your data stays within your LAN.
I’m moving to self host all my streaming stuff. Switching from local-only plex to self hosting all my media (spotify, google photos, LLMs) and tools behind a reverse proxy so i can access outside my home. It’s pretty sweet and a good learning experience using reverse proxies
Edit: Plus fuck these technofeudal lords who enclose access to markets, information, and culture.
[email protected] to the win!
Love to see it.
Edit: replied to the wrong comment
A freedom enjoyer detected. A Proper authority has been notified.
You can’t practically self-host Deepseek R1.
Look, I use the 32B distil on my 3090 every day, but it is not the same thing as full R1. And people need to stop conflating the two.
And (theoretically) API usage through one of many R1 providers is private.
Do you know of a provider is actually private? The few privacy policies I checked all had something like “We might keep some of your data for some time for anti-abuse or other reasons”…
I mean, not with certainty. If the risk of your input leaking is that great, you can just host your own VM with the 32B to be more certain.
Trust me bro, they are private
I dont really use LLMs so I didn’t even realize there were versions with different weights and stuff. I was using 7b, but found it pretty useless. Pretty sure I’m not going to be able run 32B on my rig. lmao.
guess ill continue being an LLMless pleb.
There are plenty of free LLM APIs you can use with something like Open Web UI, on any machine. I still use them myself.
Have you got any recs? I’ve got a 3080 in my machine atm
I’m not @[email protected] However here’s a pretty barebones how to article to get you started. Just know it can be as complicated as you like. For starters you may want to stick to the 7b and 14b models like mistral:7b and phi4:14b as they’ll fit easily on your card and will allow you to test the waters.
If you’re on Windows https://doncharisma.org/2024/11/23/self-hosting-ollama-with-open-webui-on-windows-a-step-by-step-guide/
If you’re using Linux https://linuxtldr.com/setup-ollama-and-open-webui-on-linux/
If you want a container https://github.com/open-webui/open-webui/blob/main/docker-compose.yaml
Locally? Arcee 14B and the 14B Deepseek distill are currently the best models that fill fit.
I’d recommend hosting them with TabbyAPI instead of ollama, as they will be much faster and more VRAM efficient. But this is more fuss.
Honestly, I would just try free APIs like Gemini, Groq, and such through open web ui, or use really cheap APIs like openrouter. Newer 14B models are okay, but they’re definitely lacking that “encyclopedic intelligence” larger models have.
I use 32b and the 672b side by side. The performance hit is around 20% and I keep all my data local. I am not conflating the two however self hosting works for me just fine. Your usecase is your own certainly. However I’d rather take the performance hit for the added data privacy.
Also it’s nice to he able to set my own weights and further distil R1
I have a local python expert a local golang expert and both have my local gitlab repository and I’ve tied their respective Ollama keys to my VSCode IDE.
Depends for sure. I usually try the 32B first, but give really “hard” queries to some API model.
With the distilled models I have, I’ve been able to build and troubleshoot pretty complicated apps in Golang and Python. However, these distilled models are very specialized and will not do things like write me a story about a duck made out of duct tape or properly summarize articles. There are absolutely limits to my workflow and setup. But I’m pretty happy with it.
Have you had any luck importing even a medium-sized codebase and doing reasoning on it? All of my experiments start to show subtle errors past 2k tokens, and at 5k tokens the errors become significant. Any attempt to ingest and process a decent-sized project (say 20k SLOC plus tooling/overhead/config) has been useless, even on models that “should” have a good-enough sized context window.
My codebase is almost 1.2GB of raw python and go files no images. I think it’s somewhere near 15k tokens for the python codebase and 22k for golang due to all the .mod and .io connectors to python libraries… it was a much bigger mess before if you can believe it.
What size model are you using? I’m getting pretty good results with R1 32b but these have been distilled to be experts in the languages of the codebases. I’m not using any general models for this.
Also it depends on the language you’re targeting as well. Rust or Lisp have issues due to how much less they’ve been documented. I think golf type languages like brainfuck are impossible. It really comes down to how the language has been documented. Python gave me issues in the beginning until I specified 3.11 in my weights and distillation/training, and that definitely fixed a lot of the hallucinations I was getting from the model.
I think static typing languages that have consistent documentation would be the easiest for this. Now that I think of it, maybe getting a typescript expert would be something I could tool around with.
Edited for legibility and the fact that I just went and looked at my datasets again. Much bigger than I initially thought.