- Author
- Posted by kashifr, with replies from many commenters
- Published
- 8 July 2025
- Kind
- Hacker News discussion thread
- Original
- news.ycombinator.com/item?id=44501413
- Subject
- Language models and training
- Themes
- Small means, Given away
What it is
The user kashifr posted the link to the Hugging Face blog post titled Smollm3. The post presents a 3B multilingual reasoning model. Commenters test it, argue about its cost and compare it with other small models.
What it says
- Commenter gardnr says the 3B model scores well and ships its full training recipe.
- sigmoid10 says reasoning mode passed 10GB of VRAM within about 1000 tokens. wizee replies that flash attention and KV cache quantization help.
- Commenters dispute the training cost of 384 H100 GPUs over 24 days. One estimate is $663,552 at $3 an hour.
- Commenters such as thatjoeoverthr advise retrieval over fine-tuning to add knowledge.
- simonw could not run it on a Mac with Ollama or llama.cpp. Commenter reach-vb says the llama.cpp change had not yet landed.
- Commenter lewtun says the team used offline preference methods, not reinforcement learning, on small models.
- Commenter refulgentis says the benchmarks omit Phi-4-mini, and SmolLM3 seemed behind it.
What evidence it gives
The thread is anecdote, personal tests and napkin cost math. Several commenters ran the model and report memory use and errors. Cost figures rest on GPU rental prices that commenters cite. Only one commenter reports checking published benchmarks, in about 10 minutes.
Interests and limits
Hacker News commenters are self-selected and mostly technical. The linked Hugging Face post was not read, only the thread. Claims about the model come from commenters, who sometimes correct each other. At least one commenter says they are from Hugging Face. Replies mix phones, laptops and servers, so memory results do not compare.