RedHatAI/SmolLM3-3B-FP8-dynamic

Model Overview

• Model Architecture: SmolLM3-3B
  • Input: Text
  • Output: Text
• Model Optimizations:
  • Weight quantization: FP8
  • Activation quantization: FP8
• Release Date: 07/28/2025
• Version: 1.0
• License(s): Apache-2.0
• Model Developers: RedHat (Neural Magic)

Model Optimizations

This model was obtained by quantizing activation and weights of SmolLM3-3B to FP8 data type. This optimization reduces the number of bits used to represent weights and activations from 16 to 8, reducing GPU memory requirements (by approximately 50%) and increasing matrix-multiply compute throughput (by approximately 2x). Weight quantization also reduces disk size requirements by approximately 50%.

Only weights and activations of the linear operators within transformers blocks are quantized. Weights are quantized with a symmetric static per-channel scheme, whereas activations are quantized with a symmetric dynamic per-token scheme. The llm-compressor library is used for quantization.

Deployment

This model can be deployed efficiently using the vLLM backend, as shown in the example below.

from vllm import LLM, SamplingParams
from transformers import AutoTokenizer

model_id = "RedHatAI/SmolLM3-3B-FP8-dynamic"
number_gpus = 1

sampling_params = SamplingParams(temperature=0.7, top_p=0.8, max_tokens=256)

tokenizer = AutoTokenizer.from_pretrained(model_id)

messages = [
    {"role": "system", "content": "You are a pirate chatbot who always responds in pirate speak!"},
    {"role": "user", "content": "Who are you?"},
]

prompts = tokenizer.apply_chat_template(messages, tokenize=False)

llm = LLM(model=model_id, tensor_parallel_size=number_gpus)

outputs = llm.generate(prompts, sampling_params)

generated_text = outputs[0].outputs[0].text
print(generated_text)

vLLM aslo supports OpenAI-compatible serving. See the documentation for more details.

Creation

Creation details

This model was created with [llm-compressor](https://github.com/vllm-project/llm-compressor) by running the code snippet below.

from transformers import AutoModelForCausalLM, AutoTokenizer
from llmcompressor.modifiers.quantization import QuantizationModifier
from llmcompressor.transformers import oneshot

# Load model
model_stub = "HuggingFaceTB/SmolLM3-3B"
model_name = model_stub.split("/")[-1]

tokenizer = AutoTokenizer.from_pretrained(model_stub)

model = AutoModelForCausalLM.from_pretrained(
    model_stub,
    device_map="auto",
    torch_dtype="auto",
)

# Configure the quantization algorithm and scheme
recipe = QuantizationModifier(
    targets="Linear",
    scheme="FP8_dynamic",
    ignore=["lm_head"],
)

# Apply quantization
oneshot(
    model=model,
    recipe=recipe,
)

# Save to disk in compressed-tensors format
save_path = model_name + "-FP8-dynamic"
model.save_pretrained(save_path)
tokenizer.save_pretrained(save_path)
print(f"Model and tokenizer saved to: {save_path}")

Evaluation

This model was evaluated on the well-known reasoning tasks: AIME24, Math-500, and GPQA-Diamond. In all cases, model outputs were generated with the vLLM engine, and evals are collected through LightEval library.

Evaluation details

  export VLLM_WORKER_MULTIPROC_METHOD=spawn
  export MODEL="RedHatAI/SmolLM3-3B-FP8-dynamic"
  export MODEL_ARGS="model_name=$MODEL,dtype=auto,max_model_length=65536,gpu_memory_utilization=0.9,tensor_parallel_size=1,add_special_tokens=False,generation_parameters={max_new_tokens:32768,temperature:0.6,top_p:0.95}"

  export TASK=aime24 # {aime24, math_500, gpqa:diamond}

  lighteval vllm $MODEL_ARGS "lighteval|${TASK}|0|0" \
      --use-chat-template \
      --output-dir out_dir

Accuracy

Category	Benchmark	HuggingFaceTB/SmolLM3-3B	RedHatAI/SmolLM3-3B-FP8-dynamic (this model)	Recovery
Reasoning	AIME24 (pass@1:64)	45.31	47.50	104.83%
	MATH-500 (pass@1:4)	89.30	88.30	98.88%
	GPQA-Diamond (pass@1:8)	41.22	40.91	99.25%
	Average	58.61	58.90	100.5%