I find that on my M2 Mac that number is a rough approximation to how much memory the model needs (usually plus about 10%) - which matters because I want to know how much RAM I will have left for running other applications.
Anything below 20GB tends not to interfere with the other stuff I'm running too much. This model looks promising!
The models feel pretty snappy when interacting with them directly via ollama, not sure about the TPS
However I've also ran into 2 things: 1) most models don't support tools, sometimes it's hard to find a version of the model that correctly uses tools, 2) even with good TPS, since the agents are usually doing chain-of-thought and running multiple chained prompts, the experience feels slow - this is even true with Cursor using their models/apis
In ollama, how do you set up the larger context, and figure out what settings to use? I've yet to find a good guide. I'm also not quite sure how I should figure out what those settings should be for each model.
There's context length, but then, how does that relate to input length and output length? Should I just make the numbers match? 32k is 32k? Any pointers?
Ollama breaks for me. If I manually set the context higher. The next api call from clone resets it back.
And ollama keeps taking it out of memory every 4 minutes.
LM studio with MLX on Mac is performing perfectly and I can keep it in my ram indefinitely.
Ollama keep alive is broken as a new rest api call resets it after. I’m surprised it’s this glitched with longer running calls and custom context length.
The SWE-Bench scores are very, very high for an open source model of this size. 46.8% is better than o3-mini (with Agentless-lite) and Claude 3.6 (with AutoCodeRover), but it is a little lower than Claude 3.6 with Anthropic's proprietary scaffold. And considering you can run this for almost free, this is a very extraordinary model.
No. I am referring to Claude 3.5 Sonnet New, released October 22, 2024, with model ID claude-3-5-sonnet-20241022, colloquially referred to as Claude 3.6 Sonnet because of Anthropic's confusing naming.
It's the reason Anthropic called their next release 3.7 Sonnet - the 3.6 version number was already being used by some in the community to refer to their 3.5v2.
What hardware are y'all using when you run these things locally? I was thinking of pre ordering the Framework desktop[0] for this purpose, but I wouldn't mind having a decent laptop that could run it (ideally Linux).
The same page also gives instructions for running the model through VLLM on a GPU, but it doesn't seem like it supports quantization, so it may require multiple GPUs since the instructions say "with at least 2 GPUs".
Does Nvidia have integrated memory options that allow you to get up to 64GB+ of VRAM without stringing together a bunch of 4090s?
For local LLMs Apple Silicon has really shown the value of shared memory, even if that comes at the cost of raw GPU power. Even if it's half the speed of an array of GPUs, being able to load the mid-sized models at all is a huge plus.
*For people without a 24GB RAM video card, I've got an 8GB RAM one running this model performs OK for simple tasks on ollama but you'd probably want to pay for an API for anything using a large context window that is time sensitive:*
Compared to an API call that finishes in about 20% of the time it feels a bit slow without the recommended graphics card and what not is all I'm saying.
In terms of benchmarks, it seems unusually well tuned for the model size but I suspect its just a case of gaming the measurement by testing against it as part of the development of the model which is not bad in and of itself since I suspect every LLM who is in this space marketed to IT folks does the same thing tbh so its objective enough given that as a rough gauge of "Is this usable?" without heavy time expense testing it.
I've worked with other models from All Hands recently, and I believe they were based on Mistral.
My general impression so far is that they aren't quite up to Claude 3.7 Sonnet, but they're quite good. More than adequate for an "AI pair coding assistant", and suitable for larger architectural work as long as you break things into steps for it.
This is basically the Mistral niche. If you are doing something generally perceived as ethical, you would use Gemma 3 IMO. When you aren't... well there are Apache licensed LLMs for you.
IMO, it's not about ethics, it's about legal risks. What if you want to fine tune a model on output related to your usage? Then my understanding is that all these derivatives need to be under the same license. What if G will change their prohibited use policy (the first line there is that they could update it from time to time)? There's really crazy stuff in terms of use of some services, what if G adds something in the same tune there which basically makes your application impossible.
It's not about ethical or not, it's about risk to your startup. Ethics are super subjective (and often change based on politics). Apache means you own your own model, period.
Maybe the EU should cover the cost of creating this agent/model, assuming it really delivers what it promises. It would allow Mistral to keep focusing on what they do and for us it would mean that the EU spent money wisely.
Wow. I was just grabbing some models and I happened to see this one while I was messing with tool support in LLamaIndex. I have an agentic coding thing I threw together and I have been trying different models on it and was looking to throw ReAct at it to bring in some models that don't have tool support and this just pops into existence!
I'm not able to get my agentic system to use this model though as it just says "I don't have the tools to do this". I tried modifying various agent prompts to explicitly say "Use foo tool to do bar" without any luck yet. All of the ToolSpec that I use are annotated etc. Pydantic objects and every other model has figured out how to use these tools.
Maybe someone here can suggest tools or at least where to look; what are the state-of-the-art models to run locally on relatively low power machines like a MacBook Air? Is there anyone tracking what is feasible given a machine spec?
"Apple Intelligence" isn't it but it would be nice to know without churning through tests whether I should bother keeping around 2-3 models for specific tasks in ollama or if their performance is marginal there's a more stable all-rounder model.
The best general model you can run locally is probably some version of Gemma 3 or the latest Mistral Small. On a Windows machine, this is limited by VRAM, since system RAM is too low-bandwidth to run models at usable speeds. On an M-series Mac, the system memory is on-die and fast enough to use. What you can run will be the total RAM, minus whatever MacOS uses and the space you want for other programs.
To determine how much space a model needs, you look at the size of the quantized (lower precision) model on HuggingFace or wherever it's hosted. Q4_K_M is a good default. As a rough rule of thumb, this will be a little over half the size of the parameters, if they were in gigabytes. For Devstral, that's 14.3GB. You will also need 1-8GB more than that, to store the context.
For example:
A 32GB Macbook Air could use Devstral at 14.3+4GB, leaving ~14GB for the system and applications.
A 16GB Macbook Air could use Gemma 3 12B at 7.3+2GB, leaving ~7GB for everything else.
An 8GB Macbook could use Gemma 3 4B at 2.5GB+1GB, but this is probably not worth doing.
I use qwen3:30b-a3b-q4_K_M for coding support and spam filtering, qwen2.5vl:32b-q4_K_M for image recognition/tagging/describing and sometimes gemma3:27b-it-qat for writing. All through Ollama, as that provides a unified interface, and then accessed from Emacs, command-line llm tool or my Clojure programs.
There is no single "best" model yet, it seems.
That's on an M4 Max with 64GB of RAM. I wish I had gotten the 128GB model, though — given that I run large docker containers that consume ~24GB of my RAM, things can get tight.
I would recommend just trying it out! (as long as you have the disk space for a few models). llama.cpp[0] is pretty easy to download and build and has good support for M-series Macbook Airs. I usually just use LMStudio[1] though - it's got a nice and easy-to-use interface that looks like the ChatGPT or Claude webpage, and you can search for and download models from within the program. LMStudio would be the easiest way to get started and probably all you need. I use it a lot on my M2 Macbook Air and it's really handy.
It's nice that Mistral is back to releasing actual open source models. Europe needs a competitive AI company.
Also, Mistral has been killing it with their most recent models. I pay for Le Chat Pro, it's really good. Mistral Small is really good. Also building a startup with Mistral integration.
I'd use a "subpar" LLM for any coding practice where I want to do the bulk of the thinking and where I care about how much coal I'm burning.
It's kind-of like asking, for which kind of road-trip would you use a Corolla hatchback instead of a Jeep Grand Wagoneer? For me the answer would be "almost all of them", but for others that might not be the case.
from what I gather it's finetuned to use OpenHand specifically so shows value on thsoe benchmark that target a whole system as a blackbox (i.e. agent + llm) more than directly target the llm input/outputs
I feel this is part of a larger and very old business trend.
But do we need 20 companies copying each other and doing the same thing?
Like, is that really competition? I'd say competition is when you do something slightly different, but I guess it's subjective based on your interpretation of what is a commodity and what is proprietary.
To my view, everyone is outright copying and creating commodity markets:
OpenAI: The OG, the Coke of Modern AI
Claude: The first copycat, The Pepsi of Modern AI
Mistral: Euro OpenAI
DeepSeek: Chinese OpenAI
Grok/xAI: Republican OpenAI
Google/MSFT: OpenAI clone as a SaaS or Office package.
I don't think they're actually the same thing, I definitely feel like Claude is much better with code than ChatGPT is so there are clearly differences in the capabilities of these models. One analogy that I find helpful here is the idea that these AIs are like animals. Just like there are animals of the same family (meaning they're genetically related to some degree) they still adapt to different niches. I see all these AI companies ultimately creating models analogous to that.
Some AIs will be good at coding (perhaps in a particular language or ecosystem), some at analyzing information and churning out a report for you, and some will be better at operating in physical spaces.
Will check it out. I like that we are all on the same page that Openness is a numerical value rather than a boolean, the challenge now is how to measure and define it, especially with ML
The SWE bench is super impressive of model of any size. However just providing one benchmark results and having to do partnership with OpenHands seems like they focused too much on optimizing the number.
What a ripoff, considering that a 5090 with 32GB of VRAM also currently costs $3k ;)
(Source: I just received the one I ordered from Newegg a week ago for $2919. I used hotstocks.io to alert me that it was available, but I wasn’t super fast at clicking and still managed to get it. Things have cooled down a lot from the craziness of early February.)
When I needed 21 3090s and none were available but for ridiculously high prices, I bought Dell Alienware comps, stripped them out, and sold the rest. Definitely made my money back mining for crypto with those cards. Dell surprisingly has a lot of computers with great RTX cards in stock.
That's probably because the 5000 series seems to be a big let-down. It's pretty much identical to the 4000 series in efficiency; they've only increased performance by massively increasing power usage.
More like 700 € if you are lucky. Prices are still not back down from the start of the AI boom.
I am hopeful that the prices will drop a bit more with Intel's recently announced Arc Pro B60 with 24GB VRAM, which unfortunately has only half the memory bandwidth of the RTX 3090.
Not sure why other hardware makers are so slow to catch up. Apple really was years ahead of the competition with the M1 Ultra with 800 GB/s memory bandwidth.
I find that on my M2 Mac that number is a rough approximation to how much memory the model needs (usually plus about 10%) - which matters because I want to know how much RAM I will have left for running other applications.
Anything below 20GB tends not to interfere with the other stuff I'm running too much. This model looks promising!
reply