Seems to completely ignore usage of local/free models as well as anything but Sonnet/ChatGPT. So my confidence in the good faith of the author is... heavily restricted.
Also, companies host for example an Exchange server on prem; and guess, what it connects to? Why you can usually access account at outlook.com?
Mind sharing a clarification on your understanding of "common" and "big"?
I am sure MS employees need to tell themselves that to sleep well. The statement itself doesn't seem to hold much epistemological value above that though.
Absolutely there are specific companies or industries where they think the risk is too great but for many, outsourcing the process is either the same or less risk then doing it all inhouse.
A 5090 has 32GB of VRAM allowing you to run a 32B model in memory at Q6_K.
You can run larger models by splitting the GPU layers that are run in VRAM vs stored in RAM. That is slower, but still viable.
This means that you can run the Qwen3-Coder-30B-A3B model locally on a 4090 or 5090. That model is a Mixture of Experts model with 3B active parameters, so you really only need a card with 3B of VRAM so you could run it on a 3090.
The Qwen3-Coder-480B-A35B model could also be run on a 4090 or 5090 by splitting the active 35B parameters across VRAM and RAM.
Yes, it will be slower than running it in the cloud. But you can get a long way with a high-end gaming rig.
> The Qwen3-Coder-480B-A35B model could also be run on a 4090 or 5090 by splitting the active 35B parameters across VRAM and RAM.
Reminds me of running Doom when I had to hack config.sys to forage 640KB of memory.
Less than 0.1% of the people reading this are doing that. Me, I gave $20 to some cloud service and I can do whatever the hell I want from this M1 MBA in a hotel room in Japan.
As long as it's within terms and conditions of whatever agreement you made for that $20. I can run queries on my own inference setup from remote locations too
The good old days of having to do crazy nutty things to get Elite II: Frontier, Magic Carpet, Worms, Xcom: UFO Enemy Unknown, Syndicate et cetera to actually run on my PC :-)
I'm finding the difference just between Sonnet 4 and Sonnet 4.5 to be meaningful in terms of the complexity of tasks I'm willing to use them for.
That doesn't mean "not plateauing".
It's better, certainly, but the difference between SOTA now and SOTA 6 months ago is a fraction of the difference between SOTA 6 months ago and the difference 18 months ago.
It doesn't mean that the models aren't getting better, it means that the improvement in each generation is smaller than the the improvement in the previous generation.
Comparing a 12 month period to a 6 month period feels unfair to me though. I think we will have a much fuller picture by the end of the year - I have high expectations for the next wave of Chinese models and for Gemini 3.
Okay. Let me clarify then.
The difference between SOTA now and SOTA 6 months ago is a fraction of the difference between SOTA 6 months ago and SOTA 12 months ago.
That still "plateauing". The performance of the models, should you take the time to chart them, is clearly asymptotic and we're in the flattening out phase now.
I also observe that all the models are converging on roughly the same performance, which makes me think that we are approaching some maxima with the current approach.
Is there a digit missing? I don't understand why this existing in 5 years is absurd
Depends almost completely on usage. No one is renting out hardware 24x7 and making a loss on it.
If you only have sporadic use then renting is better. If you're running it almost all the time of purchasing it outright is better.
In that scenario the case is even weaker for the rented-hardware model - if you're going to have a gaming rig, you're only paying a little bit more on top for a GPU with more RAM, not the full cost of the rig.
The comparison then is the extra cost of using a 24GB GPU over a standard gaming rig GPU (12GB? 8GB?) versus the cost of renting the GPU whenever you need it.
I could either spend $20 a month for my cursor license.
Or
Spend $2k+ upfront to build a machine to run models locally. Pay for the electricity cost and time to set both the machine and software up.
You said this was in the context of a gaming rig. You're not spending an extra $2k on your gaming rig to run models locally.
If you're building a dedicated LLM machine OR you're using less compute than you are paying the provider for, then, yup - $20/m is cheaper.
When you start using the model more, or if you're already building a gaming rig, then it's going to be cheaper to self-host.
So again, the economics don’t really make sense except in specific edge cases or for folks that don’t want to pay vendors. Also please don’t use italics, I don’t know why but every time you see them used it’s always a silly comment.
You can gain a lot of performance by using optimal quantization techniques for your setup(ix, awq etc), different llamacpp builds do different between each other and very different compared to something like vLLM
I spent last weekend experimenting with Ollama and LM studio. I was impressed at how good Qwen3-Coder is. Not as good as Claude, but close - maybe even better in some ways.
As I understand it, the latest Macs are good for local LLMs due to their unified memory. 32GB of RAM in one of the newer M-series seems to be the "sweet spot" for price versus performance.
There was even a recent release of Granite4 that runs on a Raspberry Pi.
https://github.com/Jewelzufo/granitepi-4-nano
For my local work I use Ollama. (M4 Max 128GB)
- gpt-oss. 20b or 120b depending on complexity of use cases.
- granite4 for speed and lower complexity (around the same as gpt20b).
Using Qwen3:32b on a 32GB M1 Pro may not be "close to cloud capabilities" but it is more than powerful enough for me, and most importantly, local and private.
As a bonus, running Asahi Linux feels like I own my Personal Computer once again.
Running smaller models on Apple Silicon is kinder on the environment/energy use and has privacy benefits for corporate use.
Using a hybrid approach makes sense for many use cases. Everyone gets to make their own decisions; for me, I like to factor in externalities like social benefit, environment, and wanting the economy to do as well as it can in our new post-mono polar world.
Inference for new releases is routinely bugged for at least a month or two as well, depending on how active the devs of a specific inference engine are and how much model creators collaborate. Personally, I hate how data from GPT's few week (and arguably somewhat ongoing) sycophancy rampage has leaked into datasets that are used for training local models, making a lot of new LLM releases insufferable to use.
Ollama, for example, will let you run any available model on just about any hardware. But using the CPU alone is _much_ slower than running it on any reasonable GPU, and obviously CPU performance varies massively too.
You can even run models that are bigger than available RAM too, but performance will be terrible.
The ideal case is to have a fast GPU and run a model that fits entirely within the GPU's memory. In these cases you might measure the model's processing speed in tens of tokens per second.
As the idealness decreases, the processing speed decreases. On a CPU only with a model that fits in RAM, you'd be maxing out in the low single digit tokens per second, and on lower performance hardware, you start talking about seconds over token instead. If the model does not fit in RAM, then the measurement is minutes per token.
For most people, their minimum acceptable performance level is in the double digit tokens per second range, which is why people optimize for that with high-end GPUs with as much memory as possible, and choose models that fit inside the GPU's RAM. But in theory you can run large models on a potato, if you're prepared to wait until next week for an answer.
> It's really just a performance tradeoff, and where your acceptable performance level is.
I am old enough to remember developers respecting the economics of running the software they create.
Ollama running locally paired occasionally with using Ollama Cloud when required is a nice option if you use it enough. I have twice signed up and paid $20/month for Ollama Cloud, love the service, but use it so rarely (because local models so often are sufficient) that I cancelled both times.
If Ollama ever implements a pay as you go API for Ollama Cloud, then I will be a long term customer. I like the business model of OpenRouter but I enjoy using Ollama Cloud more.
I am probably in the minority, but I wish subscription plans would go away and Claude Code, gemini-cli, codex, etc. would all be only available pay as you go, with ‘anti dumping’ laws applied to running unsustainable businesses.
I don’t mean to pick on OpenAI, but I think the way they fund their operations actually helps threaten the long term viability of our economy. Our government making the big all-in bet on AI dominance seems crazy to me.
If you have 8GB of RAM you can even try running them directly in Chrome via WebAssembly. Here's a demo running a model that's less than 1GB to load, entirely in your browser (and it worked for me in mobile safari just now): https://huggingface.co/spaces/cfahlgren1/Qwen-2.5-WebLLM
The Qwen model I am using is fairly small but does the job I need it to for classifying headlines pretty decently. All I ask it to do is whether a specific headline is political or not. It only responds to me with in True or False.
I access this model from an app (running locally) using the `http://localhost:11434/api/generate` REST api with `think` set to false.
Note that this qwen model is a `thinking` model. So disabling it is important. Otherwise it takes very long to respond.
Note that I tested this on my newer M4 Mac mini too and there, the performance is a LOT faster.
Also, on my new M4 Mac, I originally tried using the Apple's built in Foundation Models for this task and while it was decent, many times, it was hitting Apple's guardrails and refusing to respond because it claimed the headline was too sensitive. So I switched to the Qwen model which didn't have this problem.
Note that while this does the job I need it to, as another comment said, it won't be much help for things like coding.