Looks like ballpark a million dollars of GPU time if you want to train up one for yourself (4000 gpus/24 days).
Very nice write up that’s generous in sharing their learnings.
This is a solid and positive contribution.
Like if I really really just wanted to build it from scratch, could I do so? (not that I have that money but just curious)
It was 24 days (576 hours) not 24 hours. $663,552 @ $3/hr.
Found this a few days ago which might be neat for finding cheaper https://www.primeintellect.ai/
No affiliation with either
What's the source for $3/h?
For context, I dev an LLM client, a core tenant is keeping local as close to cloud parity as much as is possible. (via llama.cpp)
Companies aren't taking local AI seriously on a sustained basis outside Microsoft.
Overall, I usually would bite my tongue. HF is a great citizen, and I doubt this'll be a one off. However, when I see superlatives affirmed, while leaving out the local SoTA for many many moons that is a godsend in this sector, I think it is good to, rather than shy away, stand up and say this.
> We're releasing SmolLM3 with our engineering blueprint. It includes architecture details, exact data mixtures showing how we progressively boost performance across domains in a three-stage pretraining approach, and the methodology for building a hybrid reasoning model. Usually, achieving these results would require months of reverse engineering. Instead, we're providing the full methodology.
So this is primarily great for enterprises who want to do on-prem with limited budgets and maybe high-end enthusiasts.
How many tokens/s? I can't imagine that this would run in any practical way.
I've been using the smollm base models for my own finetunes just because they're so high quality, it looks like I might be using them to drive local agents/code completion in the near future too.
Their RL algorithm looks interesting. I'm still using OpenAI's algorithm for my stuff, I've been meaning to check on the SoTA since I know my code is pretty outdated (It's crazy how fast that happens with this stuff.)
./llama.cpp/llama-cli -hf unsloth/SmolLM3-3B-GGUF:Q4_K_XL --jinja -ngl 99
This seems to be a persistent issue with almost all weight releases, even from bigger companies like Meta.
Are the people who release these weights not testing them in various inference engines? Seems they make it work with Huggingface's Transformers library, then call it a day, but sometimes not even that.
My take is large labs with closed source models also did have issues during the beginning, but most likely have standardized the chat template (for eg OpenAI using ChatML). The OSS community on the other hand keeps experimenting with new templates - for example adding tool calling causes a large headache. For example in https://unsloth.ai/blog/phi3 - we found many bugs in OSS models.
The reason is simple. There isn't much money in it. llama.cpp is free and targets lower end of the hardware spectrum. Corporations will run something else, or even more likely, offload the task to contractor.
While machine learning is not my field, I've tried to finetune Mistral 7B (following their official guide and toolset) and the results did not satisfy. Had a few very specific questions from the dataset that no matter how much I've finetuned and tweaked the process it was not able to respond with correct information.
A mix of vector search + keyword search is still better at building the right question context than expecting it to learn all the information.
I've used the pretrained dataset approach. Maybe building syntethic questions and answers around the dataset yields better results but I didn't have time to experiment with that approach.
While they answer a slightly different question in the Physics of Language Models[1], based on their results it seems to me it is likely that one needs to do such augmentation of the dataset to get good results.
However, they also show that the dataset the base model is trained on can drastically affect finetuning performance. So if the base model is trained on a poor dataset for your specific task, perhaps you'll never get good performance.
[1]: https://physics.allen-zhu.com/part-3-knowledge/part-3-1
Bite the bullet and do some kind of RAG; you need to provide clear, authoritative information to a model that is skilled enough to remix it for the user.
Tuning the model to imitate the dataset will damage the model's skills and "common sense" but won't train it reliably recall information.
Also tried SmolVLM 256M and 500M, they load faster and you can embed them in assets, they work if you know what you're doing
Just keep in mind that smaller models don't perform as well due to their limited parameters
Also on Android, since you can't ship files larger than 2GB due to Java compression issues, you need to download models separately, then you can't load the model from the download folder, you have to copy it into the app's own folder, this means a Gemma 3N 2B model that's 3.14 GB would need at least 7 GB of free space on the user's phone
I hope you continue the 50-100M parameter models.
I think there is a case for models that finish fast on CPUs in solve by llm test cases.
(llama_model_load: error loading model: error loading model architecture: unknown model architecture: 'smollm3')
I've managed to run it using Python and transformers with PyTorch in device="cpu" mode but unsurprisingly that's really slow - it took 35s to respond to "say hi"!
Anyone had success with this on a Mac yet? I really want to get this running with tool calling, ideally via an OpenAI-compatible serving layer like llama-server.
The easiest would be to install llama.cpp from source: https://github.com/ggml-org/llama.cpp
If you want to avoid it, I added SmolLM3 to MLX-LM as well:
You can run it via `mlx_lm.chat --model "mlx-community/SmolLM3-3B-bf16"`
(requires the latest mlx-lm to be installed)
here's the MLX-lm PR if you're interested: https://github.com/ml-explore/mlx-lm/pull/272
similarly, llama.cpp here: https://github.com/ggml-org/llama.cpp/pull/14581
Let me know if you face any issues!
Just curious, how frequently does that happen?
My experience with phi4-mini and granite3.3 was about the same, and they annoy me even more when I hook them into code editors and try to get them to contribute to my work. For one because they're slow, and at best they suggest adding unnecessary error handling in the style of null checks everywhere, at worst they just start mixing or hallucinating programming languages. Where they would be useful as leverage if they worked, i.e. close to the edge of where I can debug and refactor without getting stuck, they just go into straight nonsense mode, especially on terse first-pass code.
Sometimes I've tried to query these things for descriptions of recent history in foreign countries, Wikipedia trivia basically, and they're very often wrong in subtle ways. For example, a politician might have been at it for half a century or so in a troubled country and because they've been ousted in a coup once in the eighties the model is absolutely sure they can't have been in office since.
If a person acted like these things do I'd wish for them to get immediate institutional care. Maybe the problem is somehow with me, but I have a deep suspicion it's not.
"So it's a small large language model?"
"Oh yes, very small."
"How can it be small and large at the same time?"
"Well, it's small by the standards of a large language model."
"So it's large."
"Oh yes, very large."
"Large compared to what?"
"Small language models."
"And so something like ChatGPT, what would that be exactly? A large large language model?"
"Yes, precisely. An LLLM."
