I usually don’t even understand the lingo they use. “Open-weighted” is the most recent one, then it usually goes down to specific “models” that everybody is supposed to know about.
These are my thoughts (I will stick to the vague “it” for now, but of course therein lies another question: “and how does all this apply to various specialised AIs”):
- Is it really feasible to run it 100% locally? I know there’s plenty of people with very powerful rigs indeed, but still. Or are 99% of these people really saying “it would, in theory, be possible to run that locally, therefore your concerns are invalid”?
- If yes to the previous: the software doesn’t come from nowhere and ultimately still relies on gas-turbine-powered datacenters and stolen IP and stolen personal data, no?
If what I wrote above is true, what exactly are people arguing when they say it’s still possible to use LLMs ethically or true to FOSS philosophy, because … ???
edit
Thanks to all who answered.
I guess it’s my fault for asking several questions in one, but this thread has attracted exactly the type of people I’m writing about; several even used the term “open-weighted models” without explaining it.
Asking to get arguments explained, I got more arguments instead.
Is it really feasible to run it 100% locally? I know there’s plenty of people with very powerful rigs indeed, but still.
There are small models you can run 100% locally on a mediocre computer or even a phone, and I mean you could be 100% offline and it will still work.
The better local models require a higher end gaming GPU or an ARM Mac with a decent amount of RAM, but nothing too extreme. You could get a computer to run them for about $2000 to $3000.
the software doesn’t come from nowhere and ultimately still relies on gas-turbine-powered datacenters and stolen IP and stolen personal data, no?
The small local models typically start with a full infustrial-size model and then they “distill” it to smaller models. The concerns over training data are still valid. I’m not sure how much the concerns over power usage are about training vs running the full industrial size models commercially. Running a small local model uses a tiny fraction of the power it takes to run an industrial size model, and the training is a one-time cost (except these companies never stop training because they want to make next year’s model better and faster).
I do think it’s worth pointing out that running a model locally keeps your conversations with it private. Remember, it can be done 100% offline. So your data won’t be stolen.
I have a friend who runs his AI locally because he doesn’t want to pay for tokens.
They are all trained on copyrighted material without permission, no LLM is ethical
do you consider “piracy” like zlibrary and annas archive unethical?
If a business is using them, yes.
i agree.
in that case, wouldn’t it be similar ethically if people use these models on their own machine for personal use?
What’s not ethical is a copyright system thag enforces artificial scarcity where there is no need for it.
Piracy is not stealing, and is not inherently unethical.
Everything is copyrighted regardless of it being owned by a corporation or blogger
I don’t partially care about the former
The absolutely most generous copyright protection length is in Mexico which is 100 years plus the life of the author when created. So lets generously say a total of 160 years. Most of the rest of the world is 120 year max. Anything outside of that is in the Public Domain. Many things had a much shorter path to the Public Domain falling into it in as little as 20 years.
So lots and lots of stuff is not copyrighted.
There is Apertus which at least claims to only use permissively licensed sources and respect robots.txt opt-outs. They outline their methodology and source datasets in this document. Though I haven’t checked the actual sources myself.
Some don’t speak any “language” they are trained to “speak” and “think” in terms of election orbitals and bonding energy. They are used in pharma and materials science to work on intractable problems like superconductivity and meds for Parkinson’s.
that’s not really an LLM then, is it?
But it is AI.
true, but that’s hardly relevant as a reply to s comment which is clearly referring to LLMs and not machine learning in general
It is, cause it uses the same architecture, https://www.geeksforgeeks.org/artificial-intelligence/exploring-the-technical-architecture-behind-large-language-models/
an example: https://github.com/bowang-lab/scGPT “talks” in RNA and is incredibly accurate and able to do predictions
Human brains are all trained on copyrighted material without permission, no human brain is ethical
As an exercise so I could learn the technology, I trained a model exclusively on the Public Domain works of L Frank Baum, specifically the “Wonderful Wizard of Oz” series (did you know there are 14 books in the series just by Baum?!). The model created is not useful as a tool, and more often than not just produces English gibberish, but every now and then it can produce original coherent statements.
Again, I didn’t do this to produce a useful tool, but rather an exercise to learn how to build models.
They are all trained on copyrighted material without permission, no LLM is ethical
However, this means I can refute your statement because 100% of the input data is public domain novels. Also, I trained it on my own hardware powered 100% by solar power.
Cool that you did that. However, “Large” as in “Large Language Model” generally refers to the dataset size, and 14 books is not Large.
The model created is not useful as a tool, and more often than not just produces English gibberish, but every now and then it can produce original coherent statements.
So we created multitudes of digital byte-sized monkeys…and gave them m/billions of typewriters. . . :p
So we created multitudes of digital byte-sized monkeys…and gave them m/billions of typewriters. . . :p
You’re not far off from describing primative LLMs. In fact, the recent innovation in the last 5 years was the application of a specific technique published in a paper called “Attention Is All You Need” that the secret sauce is “attention” added via a transformer layer. This concept is what made GPT, Claude, etc the life-like responses they have today.
They call it open-weighted and not open-source because they don’t have the training material (stolen books lol) but still want to pretend they are l33t hackers not bound to BigTech.
Yes, it’s possible to use it locally with a $2000 computer (that’s for the cheapest ones) but you’ll only get a few words per second. It doesn’t matter to the vibe coders who don’t know how to code.
The software to run that is open-source but the training of the model (the “open-weight” black box) requires to destroy the environment at least once.
Last but not least the free models are obviously censored but people don’t care about censorship anymore for some reason. “Tiananmen didn’t happen? Not my problem” without understanding that more is hidden.
Anyway, no, there is nothing ethical about it.
Ollama for text and multimodal (image to text). It supports structured JSON responses you can tie un to cool flows.
ComfyUI for image/audio/video/3d model generation. Use a docker image because installing and configuring it raw is a nightmare.
Huggingface is a good spot for the models. Ollama has its own library. There are several others.
You can get basic text interactions and small images on a typical office PC. You can make nice enough stuff with a typical gaming PC. More power definitely means better results though
Is it really feasible to run it 100% locally?
Yes. My mid 7-year-old rig featuring Ryzen 5 1600X, 16 GB DDR4 RAM, and GTX 1060 6gb can run fairly intelligent models that are actually practical to use.
The software doesn’t come from nowhere
Correct. Training the model is the most resource-intensive and planet-churning part.
I’ve come round to the idea that I will not use LLMs regardless of how “ethically” they can be sourced because their outputs are harmful in a way that builds up, like a very small dose of poison.
I’m not holding this as a hard stance, just a current analysis of the technology and how it might affect my life in my circumstances. I think we’re well into a world where we need to be deciding if some technologies don’t suit our lives, because there certainly are even more harmful technologies to come.
It might seem like a romantic or artist’s approach, but with so much spiralling out of our control I’m trying to make the active choice to be more human.
In leaving progress to the machines, in letting technology go forward on its own terms and selecting from it, with what seems to us excessive caution, modesty, or restraint, the limited though completely adequate implements of their cultures, is it possible that in thus opting not to move “forward” or not only “forward,” these people did in fact succeed in living in human history, with energy, liberty, and grace?
Always Coming Home - Stone Telling Part 3 by Ursula K. Le Guin
I’ve come round to the idea that I will not use LLMs regardless of how “ethically” they can be sourced because their outputs are harmful in a way that builds up, like a very small dose of poison.
How is LLM output any more harmful than output from a human you don’t know? I would agree with you if one were to simply blindly accept anything an LLM gives you as factual. However, I am skeptical of what humans say too. Some are inaccurate out of carelessness, some out of malice. Critical thinking is the key to protection from both errand LLM output and fallible humans.
Additionally, this thread is about locally running LLMs. One of the most dangerous aspects of most large LLMs is the sycophancy where the LLM will try to tell you what you want to hear, even if it needs to creatively invent things that don’t exist or are not true. If you are not aware, running locally means you have all the controls on the model. You’re not subject to whatever settings a large hyperscaler sets up for you. This means, among other things, you can turn down the “temperature” setting, which is the lever that controls how creative an LLM is. Practically what this means is, if you set it to “0” you are allowing it no creative action. If you ask it a question it doesn’t have actual training on, it will tell you effectively “I don’t know” instead of making something up just to have an answer.
It might seem like a romantic or artist’s approach, but with so much spiralling out of our control I’m trying to make the active choice to be more human.
The older I get the more disappointed I get in humanities large group decisions and actions.
Is it really feasible to run it 100% locally?
Yes. My Macbook Air (M2) released in 2022 can run many publicly available LLM models. The ouput is not as fast as using a large powerful datacenter, but for my local needs, I’m not in a hurry. I get about 17 to 30 tokens per second speed running 100% locally.
If yes to the previous: the software doesn’t come from nowhere
For Mac users, the interface comes from here . For the specific LLM models that is a separate question for each.
and ultimately still relies on gas-turbine-powered datacenters
DCs generating power on-site is a relatively new phenomon because existing grids are at capacity, so the only way to bring new DCs online is locally generating power at that DC, usually using gas turbines or even worse, diesel generators. Most if not all of the publicly available models for running on your own hardware were built before those gas-turnbine-generating DCs were a thing.
The public models people are running now have existed for a number of years are likely made on regular utility grid power which is whatever that nation and region uses.
and stolen IP and stolen personal data, no?
The Llama LLM is made by Meta, so probably yes for that one. Deepseek is from an AI research lab in China. QWEN is from Chinese company Alibaba. We don’t know for sure the inputs that created the Chinese models. US AI companies claim a number of the Chinese models are derived from American LLMs, but I haven’t seen (or looked for) proof of these claims.
If what I wrote above is true, what exactly are people arguing when they say it’s still possible to use LLMs ethically
Likely they mean because you don’t have to pay a large LLM owner in a rent-seeking model to run LLMs, nor does a person’s use contribute to further development by those companies.
or true to FOSS philosophy, because … ???
With the open-weight models it doesn’t rely on a commercial license to use, and the interfaces can be truly open source.
OK but what does open-weight model mean?
“Weights” in LLMs are the “final answer” numbers from the results of model training, and these are the engine of the LLM model. Lets use chocolate chip cookies as an analogy.
The chocolate chip cookies are produced with ingredients, a recipe, labor effort to produce the dough, and then cooking energy/effort to bake chocolate chip cookies. In a traditional AI company, the company gets the ingredients, they write their own recipe, do all the dough creation, and then the energy/labor for baking, and charge you money to get chocolate chip cookies.
An open-weights AI company got all the ingredients, wrote their own recipe, labor effort to produce the dough, and instead of charging you for the dough, you get as much uncooked dough as you want forever. Your only task is to take the dough and cook it yourself and you’ve got free chocolate chip cookies. You can make as many cookies as you want with your own oven. However, you are not given free raw ingredients, nor are your given the recipe to alter it in a way you might like. You can only get the dough for free.
So open-weight AI models (uncooked dough) are LLMs you can use on your own computers (oven) for free and have LLM output ( chocolate chip cookies), but you don’t get the training data (ingredients) nor the training parameters (recipe) that built the open-weight model.
Thanks, that was properly eli5’d!
It means the weights (the numeric data that makes up the model) are publicly available to download and use for free.
In some cases there are conditions such as, if you run the model as your business, you need to purchase a license. That’s usually the largest, most powerful models, though, not those most people would be able to run at home.
Question, what do you do with it and how is the data that underlies the model harvested?
Question, what do you do with it
I use it for both fun throwaway stuff as well as productivity tasks for learning.
Example of fun throwaway:
Do you remember that episode of Seinfeld where the Kramer starts using Facebook marketplace and starts buying the most worthless items before being robbed when trying to get a too-good-to-be-true sale? No? Because it never happened, but you can plug that premise into an crafted LLM prompt and it will pop out a whole TV script with in-character dialog for each actor as well as use of popular existing sets.
Foreign language learning:
I’m studying a foreign language and want to interact with just the level of vocabulary and grammar I have knowledge of right now at my level for practice. I can prompt the LLM to limit itself to just what I know now and adjust the conversation level so I can practice. If I ever get stuck, I can ask the LLM to explain the grammar usage or vocabulary choice.
and how is the data that underlies the model harvested?
I’m not sure what you’re asking here. Are you asking, for example, how the Deepseek model was trained? If so, I answered that above. If not, can you rephrase your question?
Is it really feasible to run it 100% locally? I know there’s plenty of people with very powerful rigs indeed, but still. Or are 99% of these people really saying “it would, in theory, be possible to run that locally, therefore your concerns are invalid”?
Very useful 35B models need (more or less) 8GB of VRAM and 24GB+ RAM to be usable locally. 16GB RAM might work with an RTX Nvidia card and an exl3 quantized model.
I know we are in a RAM apocalypse, but before that, that’s a quite reasonable requirement, IMO.
Personally, I run Deepseek V4 07-31 Flash at 19 tokens/second on a desktop with a single RTX 3090 and 128GB CPU RAM, and that’s a frontier class model. Again, that’s expensive these days, but pre-ram apocalypse, that is not an unreasonable workstation/homelab.
If yes to the previous: the software doesn’t come from nowhere and ultimately still relies on gas-turbine-powered datacenters…
Most open weights LLMs are Chinese. And they are:
-
Trained on pennies. They have to be, as they simply do not have a sea of GPUs like Big Tech. Training costs for their large models are in the millions or tens of millions; a single steel forge has used more energy than all of those training runs combined.
-
China relies more on renewables, and I believe the datacenters aren’t so hastily constructed with gas turbines in the middle of cities.
…and stolen IP and stolen personal data?
Yep.
This is the big caveat.
You can avoid this. Nvidia Nemotron models, for example, are trained on completely open datasets you can download yourself: https://huggingface.co/nvidia
They are very good, but just behind state of the art.
But in practice, the SOTA models most run use private datasets. Lord knows where the Chinese get it from, but given some common quirks between models, at least some data sources are shared (and possibly government provided?)
…However.
I would argue providing the result of the training as Apache licensed weights counts as “fair use,” in the same way non commercial fan works do.
They aren’t making a dime off releasing those weights. I’m not trying to sell anyone anything when I use them.
Now, the Chinese LLM services they charge for? I have no excuse for that. Once money is on the table, IP theft is definitely a thing.
completely open datasets
Not completely – it is mostly open, but they use a dozen or so private datasets for things like training on global regulations, minesweeper (for some reason), etc. To their credit, they do indicate this on the model cards, but it’s not entirely clear what is in those datasets either.
(I fell for that bit of marketing myself awhile back.)
Where is the IP theft if money isn’t changing hands?
Money not changing hands is pretty much the definition of theft.
What I’m saying is it’s akin to writing a fanfic or making fanart of your favorite franchise. Or getting inspired by a painting you see, and making something similar yourself.
Do that for your personal enjoyment? That’s fair use, under the law.
But the moment you start trying to sell it is when you get in legal hot water, and when it’s indeed morally problematic.
The scale is different, but I’d argue a similar principle applies: if you use some model trained on public works from a protected IP, but the model and its outputs are not resold, nor profited from, it’s not theft. The point is beyond money not changing hands; there’s no profit being made from the original author’s stuff. They aren’t being taken advantage of any more than someone viewing their public stuff for free, or someone creating derivatives from private work.
But all that is off the table the moment profit and distribution is involved.
Unlike the definition of heft which is a photo of OP’s mom.
-
To answer a lesser answered sub question of a question: what is a weight?
Imagine you have an output, its either a or b. You have an input, whatever it is. And then there is a weight in the middle. The weight determines wether the output will be a or b.
In math it would be a function input*weight=output.
ANN (Artificial Neural Networks, just afancy name) is a chain of a whole bunch of weights. There are layers and each layer generates an output from a bunch of weights and the next layer uses it as inputs. The name comes from the visualization, which looks like neurons firing to other neurons (from biology, things that make your brain work).
I don’t know if 5 year olds are allowed to watch 1.5h videos, but if so: this one has all the pro and con arguments explained nicely and makes a point why it’s not a good idea to shame people for using LLMs.
Some companies build out more electrical capacity than they use. Substations are funded by the DC and built to twice the desired capacity - and half is delivered to the community. The generators are run only when utility power goes down, and run on diesel.
However, not all companies are this responsible and the grievously irresponsible ones get the press and make the responsible companies look bad.
Can you name how many data centers have actually built out excess capacity before they started operating?
Most of the ones I’ve heard have simply claimed that they might eventually, which is a corporate way of saying we’ll walk that back once we think no one is paying attention.
I’ll take a crack at it.
I run AI on an old Nvidia p40 purchased online for less than $200. The models I run on it came predominantly from Chinese companies and groups. On balance, they:
-
use much less fossil fuel in producing these LLMs than their Western counterparts
-
produce models that run much better on lower end consumer hardware
As for training data, the inputs used to create these models, it varies greatly but the most popular line, Qwen, Came from the company’s own data from operating such huge networks and systems for so long
I really fail to see how any of that is worse than playing a video game.
None of this takes away from the very real issues around data center build out and Western companies using the systems to scare people and continue an economic bubble. That’s all true and bad. But there’s nuance. Not all AI is created equal
use much less fossil fuel in producing these LLMs than their Western counterparts
Utterly false, given that all the decent Chinese models (especially Qwen) are distilled from Western frontier models.
They literally could not exist without the enormously carbon emissive western models existing first to train them.
yeah the comment was bad enough I looked at the user. been around for 2 years and no posts or comments till this one. made a note on the account.
Can you prove that? Alibaba (makers of Qwen) is one of the largest data processors in the world. Why do you think they’re dependent on Western systems?
They’ve been explicitly accused of doing so and it would make sense given that they don’t have access to the same level of compute as Anthropic or OpenAI, being subject to the chip embargo and all.
Ok so no proof then?
distilled from Western frontier models.
What does this mean please (remember the title of the post!) and how do you know?
Basically, it’s really really expensive (from a compute standpoint, and more compute = money and energy) to train a decent LLM just using books and conversations etc. Anthropic, OpenAI, etc spend a truly insane amount of money doing this is in data centers.
Running an LLM like that is also extremely expensive from a compute standpoint, and right now the true frontier models can basically only be run in datacenters.
However, once you have a really good LLM, you can use it to train another LLM. Because it’s basically already refined all its original training data (and is capable of further refining it’d output on the fly), training a new LLM from it is vastly easier and cheaper.
Additionally, you can use a much smaller and more efficient model, so it can run on less powerful hardware.
So all local LLMs that are any good that exist right now, were trained using the outputs of existing super expensive frontier models. They could not exist without the frontier models to train them. This is also how OpenAI and Anthropic create their cheaper models. Mythos/Fable is anthropics current frontier model and they distill it into (use it to train) their cheaper models like Sonnet and Opus.
Wow. as if the original models weren’t rickety enough.
They’re honestly not that rickety. If you’re getting your news from social media (including Lemmy) you may want to actually try them rather than just take the random incidents where things go wrong as broadly representative.
Also China has an insanely diversified grid, that utilises almost a global scale of fossil fuels.
So its like green communism at best
-
There’s a lot of info that you need to know to explore this space, so I’ll take my own shot at answering. Let me know if anything needs further explaining!
An “open weight” model is an AI model where you can download the data needed to run the model on your own hardware for free. Contrast this with proprietary models-as-a-service like ChatGPT and Claude where you have no access to the data needed to run the model yourself – you can only use it through the services provided, usually for a fee, and which can be taken away from you or changed at any time with no recourse.
The mapping to traditional open source terms does not work well since what you get is a binary artifact.
Those artifacts are released with a license – and many of the models are licensed permissively (e.g. MIT or Apache license terms). You can take those weights, modify them, and then release them as new models – and people do actually do this in practice!
Is it really feasible to run it 100% locally?
Yes. I run models on my own computers and have tried a number of configurations to figure out what works well. The Qwen family of open weight models (from Alibaba) are the ones I’ve found most useful so far. Gemma4 models (from Google) are also useful.
I prefer models that have been modified by the community to remove corporate censorship – i.e. stripping that “As a large language model…” cover-your-ass crap and evasiveness on topics like Tiananmen Square. If that means the model is technically capable of telling me to go kill myself too, so be it; I’ve spent 25+ years dealing with assholes on the internet and can handle abuse from a stupid robot if I have to. (In practice though, they’re usually pretty nice still unless I deliberately tell them to act like an asshole – and then Qwen, at least, starts to sound like a snarky redditor; it’s quite funny most of the time, actually.)
If yes to the previous: the software doesn’t come from nowhere and ultimately still relies on gas-turbine-powered datacenters and stolen IP and stolen personal data, no?
Models do require training to create, yes. It’s generally not clear where the training physically happened IRL – so, yes, some of them probably used power from gas turbines, but others may be drawing power from the Three Gorges Dam in China or solar plants or nuclear plants or whatever else is hooked up to the electric grid where the training happened. Most of them are also not very open about the data sets they were trained on. (There are exceptions to this though!) The Chinese models in particular are almost certainly trained heavily on logs extracted from Western models in addition to using whatever other data they could get ahold of. Whether you think that’s ethical or not is a matter of perspective; how do you feel about Robin Hood?
Once a model has been trained though, it can run on a normal GPU. The power requirements to run an LLM are basically the same as running a video game, or, equivalently, about the same as turning on a few incandescent lightbulbs. (The iGPU in one of my systems uses 100W; the discrete GPU in another system I’ve tried uses 215W under load with appropriate tuning – or 300W if you run it naively.)
If you want to run a model yourself, I recommend using llama.cpp – there are instructions on how to get started with it here: https://llama.app/
These are the models I’ve found most useful:
- Qwen3.8-27B (official version): https://huggingface.co/Qwen/Qwen3.8-27B
- Qwen3.8-27B (uncensored): https://huggingface.co/MuXodious/Qwen3.8-27B-absolute-heresy
- Qwen3.6-35B-A3B (official version): https://huggingface.co/Qwen/Qwen3.6-35B-A3B
- Qwen3.6-35B-A3B (uncensored – ⚠️ mildly NSFW graphics on page): https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic
- Gemma4 (official collection – many variants in different sizes): https://huggingface.co/collections/google/gemma-4
If you have an iGPU only, I recommend using one of the so called “Mixture of Experts” (MoE) releases. These are typically named like 35B-A3B or similar; the first number indicates the total number of weights (35 billion) and the second indicates how many are “active” (i.e. actually used during computation) at one time while the model is running (3 billion in the example). These models need less computation to run and stay fast on weaker GPUs. Qwen3.6-35B-A3B is very good in this space and was my go-to model for a long time.
If you have a discrete GPU and enough VRAM, I recommend using a dense model (i.e. one that activates all its weights while answering) like Qwen3.8-27B.
It’s worth noting that people don’t usually use the full quality weights (which are typically ~2 bytes per weight); they use a “quantized” version – compressed in a lossy fashion like a JPEG. Going down to 4-bits (half a byte) on average per weight is about as low as most people like to go – you will see this indicated in names like Q4_K_M. (Quantized to ~4 bits with the K quantizaation scheme, medium variant.) Usually a bigger number is better in the sense of “closer to the original quality” – at the cost of needing more RAM.
Full quality weights are often found as safetensor files on HuggingFace. Quantized weights intended for use with llama.cpp are usually in GGUF file format.
Does that help?
“Open weighted” I take to mean that the user can see what sources the LLM/AI was trained upon. Usually this information is a closely guarded corporate secret.
As for “models” this is no different than the different models that a car company offers. Sentra from Nissan, F-150 from Ford, Civic from Honda
If locally run or through a data center, LLM/AIs have two major problems:
- The results can’t really be trusted due to their tendency to hallucinate.
- Frequent LLM/AI usage has been proven to deskill the user as well as a reduction in cognitive ability. I’m dumb enough without getting help from a clanker.
Assuming “run it 100% locally” means on your own hardware with network access, I see no reason this can’t happen given a powerful enough computer.
Open weight just means you can download the model. They tried to call that open source because they’re sleazy assholes but that was successfully stopped. “Open weight” still causes confusion as you can see and that is their goal in starting that term.
