While hosted mega models make headlines for doing cool stuff, the vast majority of applications for AI simply don't need all that power, and thus cost. That’s a big part of why businesses are screaming that there’s no ROI from AI.
Brining this tech down into small local models is likely where this all converges for the vast majority of use cases and what solves the present ROI crisis for LLM-based AI.
If you are into small local models I highly recommend vibe thinker. It's a model trained specifically for reasoning. Basically a problem solver. When compared with other models, on math problems benchmarks, it's closer to models hundred times its size than ten times its size which it beats comfortably.
It supports long contexts on limited VRAM and is blazing fast.
Tangential/meta: Holy shit, I've never seen a thread where almost half the comments are dead (and LLM written), especially for a post that's (currently) at 86 points and 20 comments (4x ratio is "pretty good quality" post signal generally for me).
Those are format examples and test fixtures. Five to ten rows each. Not training data.
You did spot a real problem though. Eight configs in `examples/configs` pointed at those fixtures as training data. Seven were still on the old schema and would not even parse. I've fixed that, added a README to the folder, and a test that parses every config so it doesn't quietly break again.
Dataset size mostly depends on the task. Format or style, a few hundred examples is often enough. A task the model already half knows, usually a few thousand. New facts are often a bad fit for fine-tuning. I'd reach for RAG instead.
I haven't measured how this changes with model size, so I'm not going to make up a number.
This seems really interesting - I was curious about this line from the website.
“The whole post-training stack in one CLI. Soup doctors your data pre-flight, picks the method, writes the config, derives evals from your own data, gates every save, and self-corrects reward hacking mid-run instead of just halting.”
How does soup auto tune the hyper parameters and make some of these more complex training decisions?
Do not buy a 4 GB card for this. Mine is an RTX 3050 Laptop, I picked it because it
is boring hardware that many people already have.
If you are buying, buy VRAM. At 0.5B where I could measure both, resident training
was 1.43x faster than streaming. And if you do stream, system RAM matters more, the
base sits there and has to page-lock. 16 GB is about the floor for 8B.
Because streaming only removes the decoder stack. The embeddings and lm_head stay
resident, that is 2.10 GB of the 3.32 GB peak on 8B. And the logits tensor scales
with batch x seq x vocab, not with depth.
So it goes from "whole model must fit" to "embeddings plus one layer plus logits
must fit". That is why 8B works and why I did not try 14B.
The table on the site is the normal resident path, streaming is opt-in and BETA.
Should be clearer, my fault.
The constraint everyone works around is that the frozen base has to fit in VRAM. But during LoRA the base is frozen — read, never written. It doesn't need to live in VRAM, it needs to arrive before the matmul that uses it. So it sits in host RAM and streams into a small pool of pre-allocated VRAM buffers, one decoder layer at a time, prefetched one ahead on a dedicated CUDA stream. Peak VRAM becomes one layer instead of the whole model.
Measured on an RTX 3050 Laptop (4 GB, Windows): Llama-3.1-8B in NF4 at 119.6 tok/s, 3.32 GB peak, 100% SM occupancy. Also Qwen2.5-3B with an un-quantized bf16 base at 143 tok/s in 2.15 GB, which is CUDA OOM when trained resident on the same card. Overhead is 1.43x vs resident, measured at 0.5B — the only size on this card with a valid resident baseline, and I publish that baseline so you can check the division.
Most of the work wasn't speed, it was correctness. Streaming fails silently: cut the autograd path and the loss still falls because the upper layers keep learning. So the bar was bit-exactness against a resident reference of the same numerics — max abs logit difference 0.0, across nine architecture families in two precisions, as a CI test rather than a one-off. That protocol caught a PEFT dispatch defect producing 0.94 logit divergence with byte-identical weights and adapters, no crash, no warning.
Not claiming anything above 8B — 14B NF4 needs ~7.5 GB page-locked against a measured 7.12 GB ceiling here, so I didn't run it. All numbers are Windows, so pessimistic vs Linux.
It isn't. Kazakh and Russian.
I said this further down but that comment is dead so you would not have
seen it. The later replies are mine, written by me.
While hosted mega models make headlines for doing cool stuff, the vast majority of applications for AI simply don't need all that power, and thus cost. That’s a big part of why businesses are screaming that there’s no ROI from AI.
Brining this tech down into small local models is likely where this all converges for the vast majority of use cases and what solves the present ROI crisis for LLM-based AI.
It supports long contexts on limited VRAM and is blazing fast.
https://github.com/WeiboAI/VibeThinker
Writing them myself now.
They're all very short though. Anyone got a good rule for how much data of this nature is needed to successfully fine-tune a model of this size?
New facts is where people waste a week. The model comes back wrong in a new way. Use RAG for facts.
“The whole post-training stack in one CLI. Soup doctors your data pre-flight, picks the method, writes the config, derives evals from your own data, gates every save, and self-corrects reward hacking mid-run instead of just halting.”
How does soup auto tune the hyper parameters and make some of these more complex training decisions?
I have a couple of comments about https://trysoup.dev
> Get Started for Free
Does this mean that this will not be free at some point?
The website is difficult to read (gray on black doesn't work well for me).
So it goes from "whole model must fit" to "embeddings plus one layer plus logits must fit". That is why 8B works and why I did not try 14B.
The table on the site is the normal resident path, streaming is opt-in and BETA. Should be clearer, my fault.
The constraint everyone works around is that the frozen base has to fit in VRAM. But during LoRA the base is frozen — read, never written. It doesn't need to live in VRAM, it needs to arrive before the matmul that uses it. So it sits in host RAM and streams into a small pool of pre-allocated VRAM buffers, one decoder layer at a time, prefetched one ahead on a dedicated CUDA stream. Peak VRAM becomes one layer instead of the whole model.
Measured on an RTX 3050 Laptop (4 GB, Windows): Llama-3.1-8B in NF4 at 119.6 tok/s, 3.32 GB peak, 100% SM occupancy. Also Qwen2.5-3B with an un-quantized bf16 base at 143 tok/s in 2.15 GB, which is CUDA OOM when trained resident on the same card. Overhead is 1.43x vs resident, measured at 0.5B — the only size on this card with a valid resident baseline, and I publish that baseline so you can check the division.
Most of the work wasn't speed, it was correctness. Streaming fails silently: cut the autograd path and the loss still falls because the upper layers keep learning. So the bar was bit-exactness against a resident reference of the same numerics — max abs logit difference 0.0, across nine architecture families in two precisions, as a CI test rather than a one-off. That protocol caught a PEFT dispatch defect producing 0.94 logit divergence with byte-identical weights and adapters, no crash, no warning.
Not claiming anything above 8B — 14B NF4 needs ~7.5 GB page-locked against a measured 7.12 GB ceiling here, so I didn't run it. All numbers are Windows, so pessimistic vs Linux.
Measurement records, including the ones I threw away: https://github.com/MakazhanAlpamys/Soup/tree/main/benchmarks
Write-up: https://doi.org/10.5281/zenodo.21771064
Happy to answer anything about the scheduler or the correctness protocol.