xv0y5ncu/SmolLM2-360M-Instruct-GLQ-trellis-3inst-3bpw

SmolLM2-360M-Instruct — GLQ trellis (3INST) 3 bpw

SmolLM2-360M-Instruct quantized with GLQ using the QTIP-style trellis (TCQ) codebook at 3 bits/weight, lookup-free 3INST variant, uniform rate across all layers.

pip install glq vllm
vllm serve xv0y5ncu/SmolLM2-360M-Instruct-GLQ-trellis-3inst-3bpw --quantization glq

The ladder

All four rungs were quantized, smoke-tested and measured in one session on one RTX PRO 6000 Blackwell (vLLM 0.27.1, glq 0.8.8, full CUDA graphs, 256-token decode, chat-template sample):

rung	weights loaded	wikitext-2 PPL	avg weight SQNR	B=1 tok/s	B=32 aggregate tok/s
bf16 (reference)	—	12.735	—	—	—
6 bpw	0.31 GiB	12.755 (+0.16%)	33.20 dB	264	4,466
5 bpw	0.27 GiB	12.834 (+0.78%)	27.67 dB	264	4,383
4 bpw	0.24 GiB	13.085 (+2.7%)	22.04 dB	303	6,478
3 bpw	0.20 GiB	14.173 (+11.3%)	16.19 dB	294	6,650

TTFT was 16 ms at B=1 on every rung. The lower rungs decode faster (less memory traffic; 5-6 bpw also pay a second stacked-RVQ stage). Every rung produced coherent chat-template output in the smoke test.

Perplexity is teacher-forced wikitext-2 (seqlen 2048, 128 non-overlapping windows), measured through vLLM, with the bf16 reference run the same way in the same session — so the deltas are internally comparable, but do not mix these values into HF-measured PPL series. 6 bpw's +0.02 is within the measurement's character for bf16 itself; 3 bpw pays a real ~11%. No task evals were run at this model size — this table is the trade-off: footprint and speed against perplexity.

How it was produced

GLQ_TRELLIS_VARIANT=3inst glq-quantize \
    --model HuggingFaceTB/SmolLM2-360M-Instruct \
    --codebook trellis --bpw 3 --nsamples 128 --seqlen 2048

(trellis and 3inst are the defaults since glq 0.8.8; kept explicit for reproducibility.)

Runtime support

Validated on vLLM with glq >= 0.8.8 — the numbers above. The HF transformers path is untested for this checkpoint.

---

Original model card: SmolLM2-360M-Instruct

SmolLM2

Table of Contents

1. Model Summary
2. Limitations
3. Training
4. License
5. Citation

Model Summary

SmolLM2 is a family of compact language models available in three size: 135M, 360M, and 1.7B parameters. They are capable of solving a wide range of tasks while being lightweight enough to run on-device. More details in our paper: https://arxiv.org/abs/2502.02737

SmolLM2 demonstrates significant advances over its predecessor SmolLM1, particularly in instruction following, knowledge, reasoning. The 360M model was trained on 4 trillion tokens using a diverse dataset combination: FineWeb-Edu, DCLM, The Stack, along with new filtered datasets we curated and will release soon. We developed the instruct version through supervised fine-tuning (SFT) using a combination of public datasets and our own curated datasets. We then applied Direct Preference Optimization (DPO) using UltraFeedback.

The instruct model additionally supports tasks such as text rewriting, summarization and function calling (for the 1.7B) thanks to datasets developed by Argilla such as Synth-APIGen-v0.1. You can find the SFT dataset here: https://huggingface.co/datasets/HuggingFaceTB/smol-smoltalk and finetuning code in the alignement handbook

For more details refer to: https://github.com/huggingface/smollm. You will find pre-training, post-training, evaluation and local inference code.

How to use

Transformers

pip install transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
checkpoint = "HuggingFaceTB/SmolLM2-360M-Instruct"

device = "cuda" # for GPU usage or "cpu" for CPU usage
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
# for multiple GPUs install accelerate and do `model = AutoModelForCausalLM.from_pretrained(checkpoint, device_map="auto")`
model = AutoModelForCausalLM.from_pretrained(checkpoint).to(device)

messages = [{"role": "user", "content": "What is the capital of France."}]
input_text=tokenizer.apply_chat_template(messages, tokenize=False)
print(input_text)
inputs = tokenizer.encode(input_text, return_tensors="pt").to(device)
outputs = model.generate(inputs, max_new_tokens=50, temperature=0.2, top_p=0.9, do_sample=True)
print(tokenizer.decode(outputs[0]))

Chat in TRL

You can also use the TRL CLI to chat with the model from the terminal:

pip install trl
trl chat --model_name_or_path HuggingFaceTB/SmolLM2-360M-Instruct --device cpu

Transformers.js

npm i @huggingface/transformers

import { pipeline } from "@huggingface/transformers";

// Create a text generation pipeline
const generator = await pipeline(
  "text-generation",
  "HuggingFaceTB/SmolLM2-135M-Instruct",
);

// Define the list of messages
const messages = [
  { role: "system", content: "You are a helpful assistant." },
  { role: "user", content: "What is the capital of France?" },
];

// Generate a response
const output = await generator(messages, { max_new_tokens: 128 });
console.log(output[0].generated_text.at(-1).content);
// "The capital of France is Paris."

Evaluation

In this section, we report the evaluation results of SmolLM2. All evaluations are zero-shot unless stated otherwise, and we use lighteval to run them.

Base Pre-Trained Model

Metrics	SmolLM2-360M	Qwen2.5-0.5B	SmolLM-360M
HellaSwag	54.5	51.2	51.8
ARC (Average)	53.0	45.4	50.1
PIQA	71.7	69.9	71.6
MMLU (cloze)	35.8	33.7	34.4
CommonsenseQA	38.0	31.6	35.3
TriviaQA	16.9	4.3	9.1
Winogrande	52.5	54.1	52.8
OpenBookQA	37.4	37.4	37.2
GSM8K (5-shot)	3.2	33.4	1.6

Instruction Model

Metric	SmolLM2-360M-Instruct	Qwen2.5-0.5B-Instruct	SmolLM-360M-Instruct
IFEval (Average prompt/inst)	41.0	31.6	19.8
MT-Bench	3.66	4.16	3.37
HellaSwag	52.1	48.0	47.9
ARC (Average)	43.7	37.3	38.8
PIQA	70.8	67.2	69.4
MMLU (cloze)	32.8	31.7	30.6
BBH (3-shot)	27.3	30.7	24.4
GSM8K (5-shot)	7.43	26.8	1.36

Limitations

SmolLM2 models primarily understand and generate content in English. They can produce text on a variety of topics, but the generated content may not always be factually accurate, logically consistent, or free from biases present in the training data. These models should be used as assistive tools rather than definitive sources of information. Users should always verify important information and critically evaluate any generated content.

Training

Model

• Architecture: Transformer decoder
• Pretraining tokens: 4T
• Precision: bfloat16

Hardware

• GPUs: 64 H100

Software

• Training Framework: nanotron

License

Apache 2.0

Citation

@misc{allal2025smollm2smolgoesbig,
      title={SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model}, 
      author={Loubna Ben Allal and Anton Lozhkov and Elie Bakouch and Gabriel Martín Blázquez and Guilherme Penedo and Lewis Tunstall and Andrés Marafioti and Hynek Kydlíček and Agustín Piqueres Lajarín and Vaibhav Srivastav and Joshua Lochner and Caleb Fahlgren and Xuan-Son Nguyen and Clémentine Fourrier and Ben Burtenshaw and Hugo Larcher and Haojun Zhao and Cyril Zakka and Mathieu Morlon and Colin Raffel and Leandro von Werra and Thomas Wolf},
      year={2025},
      eprint={2502.02737},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.02737}, 
}