Ju214/Mistral-Small-24B-3.1

Mistral-Small-3.1-24B-Instruct-2503-quantized.w4a16

Model Overview

• Model Architecture: Mistral3ForConditionalGeneration
  • Input: Text / Image
  • Output: Text
• Model Optimizations:
  • Weight quantization: INT4
• Intended Use Cases: It is ideal for:
  • Fast-response conversational agents.
  • Low-latency function calling.
  • Subject matter experts via fine-tuning.
  • Local inference for hobbyists and organizations handling sensitive data.
  • Programming and math reasoning.
  • Long document understanding.
  • Visual understanding.
• Out-of-scope: Use in any manner that violates applicable laws or regulations (including trade compliance laws). Use in languages not officially supported by the model.
• Release Date: 04/15/2025
• Version: 1.0
• Model Developers: Ju214 (Neural Magic)

Model Optimizations

This model was obtained by quantizing the weights of Mistral-Small-3.1-24B-Instruct-2503 to INT4 data type. This optimization reduces the number of bits per parameter from 16 to 4, reducing the disk size and GPU memory requirements by approximately 75%.

Only the weights of the linear operators within transformers blocks are quantized. Weights are quantized using a symmetric per-group scheme, with group size 128. The GPTQ algorithm is applied for quantization, as implemented in the llm-compressor library.

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 AutoProcessor

model_id = "Ju214/Mistral-Small-24B-3.1"
number_gpus = 1

sampling_params = SamplingParams(temperature=0.7, top_p=0.8, max_tokens=256)
processor = AutoProcessor.from_pretrained(model_id)

messages = [{"role": "user", "content": "Give me a introduction to large language model."}]

prompts = processor.apply_chat_template(messages, add_generation_prompt=True, 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 AutoProcessor
from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.transformers import oneshot
from llmcompressor.transformers.tracing import TraceableMistral3ForConditionalGeneration
from datasets import load_dataset, interleave_datasets
from PIL import Image
import io

# Load model
model_stub = "mistralai/Mistral-Small-3.1-24B-Instruct-2503"
model_name = model_stub.split("/")[-1]

num_text_samples = 1024
num_vision_samples = 1024
max_seq_len = 8192

processor = AutoProcessor.from_pretrained(model_stub)

model = TraceableMistral3ForConditionalGeneration.from_pretrained(
    model_stub,
    device_map="auto",
    torch_dtype="auto",
)

# Text-only data subset
def preprocess_text(example):
    input = {
        "text": processor.apply_chat_template(
            example["messages"],
            add_generation_prompt=False,
        ),
        "images": None,
    }
    tokenized_input = processor(**input, max_length=max_seq_len, truncation=True)
    tokenized_input["pixel_values"] = tokenized_input.get("pixel_values", None)
    tokenized_input["image_sizes"] = tokenized_input.get("image_sizes", None)
    return tokenized_input

dst = load_dataset("neuralmagic/calibration", name="LLM", split="train").select(range(num_text_samples))
dst = dst.map(preprocess_text, remove_columns=dst.column_names)

# Text + vision data subset
def preprocess_vision(example):
    messages = []
    image = None
    for message in example["messages"]:
        message_content = []
        for content in message["content"]:
            if content["type"] == "text":
                message_content.append({"type": "text", "text": content["text"]})
            else:
                message_content.append({"type": "image"})
                image = Image.open(io.BytesIO(content["image"]))

        messages.append(
            {
                "role": message["role"],
                "content": message_content,
            }
        )

    input = {
        "text": processor.apply_chat_template(
            messages,
            add_generation_prompt=False,
        ),
        "images": image,
    }
    tokenized_input = processor(**input, max_length=max_seq_len, truncation=True)
    tokenized_input["pixel_values"] = tokenized_input.get("pixel_values", None)
    tokenized_input["image_sizes"] = tokenized_input.get("image_sizes", None)
    return tokenized_input

dsv = load_dataset("neuralmagic/calibration", name="VLM", split="train").select(range(num_vision_samples))
dsv = dsv.map(preprocess_vision, remove_columns=dsv.column_names)

# Interleave subsets
ds = interleave_datasets((dsv, dst))

# Configure the quantization algorithm and scheme
recipe = GPTQModifier(
    ignore=["language_model.lm_head", "re:vision_tower.*", "re:multi_modal_projector.*"],
    sequential_targets=["MistralDecoderLayer"],
    dampening_frac=0.01,
    targets="Linear",
    scheme="W4A16",
)

# Define data collator
def data_collator(batch):
    import torch
    assert len(batch) == 1
    collated = {}
    for k, v in batch[0].items():
        if v is None:
            continue
        if k == "input_ids":
            collated[k] = torch.LongTensor(v)
        elif k == "pixel_values":
            collated[k] = torch.tensor(v, dtype=torch.bfloat16)
        else:
            collated[k] = torch.tensor(v)
    return collated


# Apply quantization
oneshot(
    model=model,
    dataset=ds, 
    recipe=recipe,
    max_seq_length=max_seq_len,
    data_collator=data_collator,
    num_calibration_samples=num_text_samples + num_vision_samples,
)

# Save to disk in compressed-tensors format
save_path = model_name + "-quantized.w4a16"
model.save_pretrained(save_path)
processor.save_pretrained(save_path)
print(f"Model and tokenizer saved to: {save_path}")

Evaluation

The model was evaluated on the OpenLLM leaderboard tasks (version 1), MMLU-pro, GPQA, HumanEval and MBPP. Non-coding tasks were evaluated with lm-evaluation-harness, whereas coding tasks were evaluated with a fork of evalplus. vLLM is used as the engine in all cases.

Evaluation details

MMLU

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Mistral-Small-3.1-24B-Instruct-2503-quantized.w4a16",dtype=auto,gpu_memory_utilization=0.5,max_model_len=8192,enable_chunk_prefill=True,tensor_parallel_size=2 \
  --tasks mmlu \
  --num_fewshot 5 \
  --apply_chat_template\
  --fewshot_as_multiturn \
  --batch_size auto

ARC Challenge

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Mistral-Small-3.1-24B-Instruct-2503-quantized.w4a16",dtype=auto,gpu_memory_utilization=0.5,max_model_len=8192,enable_chunk_prefill=True,tensor_parallel_size=2 \
  --tasks arc_challenge \
  --num_fewshot 25 \
  --apply_chat_template\
  --fewshot_as_multiturn \
  --batch_size auto

GSM8k

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Mistral-Small-3.1-24B-Instruct-2503-quantized.w4a16",dtype=auto,gpu_memory_utilization=0.9,max_model_len=8192,enable_chunk_prefill=True,tensor_parallel_size=2 \
  --tasks gsm8k \
  --num_fewshot 8 \
  --apply_chat_template\
  --fewshot_as_multiturn \
  --batch_size auto

Hellaswag

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Mistral-Small-3.1-24B-Instruct-2503-quantized.w4a16",dtype=auto,gpu_memory_utilization=0.5,max_model_len=8192,enable_chunk_prefill=True,tensor_parallel_size=2 \
  --tasks hellaswag \
  --num_fewshot 10 \
  --apply_chat_template\
  --fewshot_as_multiturn \
  --batch_size auto

Winogrande

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Mistral-Small-3.1-24B-Instruct-2503-quantized.w4a16",dtype=auto,gpu_memory_utilization=0.5,max_model_len=8192,enable_chunk_prefill=True,tensor_parallel_size=2 \
  --tasks winogrande \
  --num_fewshot 5 \
  --apply_chat_template\
  --fewshot_as_multiturn \
  --batch_size auto

TruthfulQA

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Mistral-Small-3.1-24B-Instruct-2503-quantized.w4a16",dtype=auto,gpu_memory_utilization=0.5,max_model_len=8192,enable_chunk_prefill=True,tensor_parallel_size=2 \
  --tasks truthfulqa \
  --num_fewshot 0 \
  --apply_chat_template\
  --batch_size auto

MMLU-pro

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Mistral-Small-3.1-24B-Instruct-2503-quantized.w4a16",dtype=auto,gpu_memory_utilization=0.5,max_model_len=8192,enable_chunk_prefill=True,tensor_parallel_size=2 \
  --tasks mmlu_pro \
  --num_fewshot 5 \
  --apply_chat_template\
  --fewshot_as_multiturn \
  --batch_size auto

MMMU

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Mistral-Small-3.1-24B-Instruct-2503-quantized.w4a16",dtype=auto,gpu_memory_utilization=0.9,max_images=8,enable_chunk_prefill=True,tensor_parallel_size=2 \
  --tasks mmmu_val \
  --apply_chat_template\
  --batch_size auto

ChartQA

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Mistral-Small-3.1-24B-Instruct-2503-quantized.w4a16",dtype=auto,gpu_memory_utilization=0.9,max_images=8,enable_chunk_prefill=True,tensor_parallel_size=2 \
  --tasks chartqa \
  --apply_chat_template\
  --batch_size auto

Coding

The commands below can be used for mbpp by simply replacing the dataset name.

Generation

python3 codegen/generate.py \
  --model RedHatAI/Mistral-Small-3.1-24B-Instruct-2503-quantized.w4a16 \
  --bs 16 \
  --temperature 0.2 \
  --n_samples 50 \
  --root "." \
  --dataset humaneval

Sanitization

python3 evalplus/sanitize.py \
  humaneval/RedHatAI--Mistral-Small-3.1-24B-Instruct-2503-quantized.w4a16_vllm_temp_0.2

Evaluation

evalplus.evaluate \
  --dataset humaneval \
  --samples humaneval/RedHatAI--Mistral-Small-3.1-24B-Instruct-2503-quantized.w4a16_vllm_temp_0.2-sanitized

Accuracy

Category	Benchmark	Mistral-Small-3.1-24B-Instruct-2503	Mistral-Small-3.1-24B-Instruct-2503-quantized.w4a16 (this model)	Recovery
OpenLLM v1	MMLU (5-shot)	80.67	79.74	98.9%
	ARC Challenge (25-shot)	72.78	72.18	99.2%
	GSM-8K (5-shot, strict-match)	58.68	59.59	101.6%
	Hellaswag (10-shot)	83.70	83.25	99.5%
	Winogrande (5-shot)	83.74	83.43	99.6%
	TruthfulQA (0-shot, mc2)	70.62	69.56	98.5%
	Average	75.03	74.63	99.5%
	MMLU-Pro (5-shot)	67.25	66.56	99.0%
	GPQA CoT main (5-shot)	42.63	47.10	110.5%
	GPQA CoT diamond (5-shot)	45.96	44.95	97.80%
Coding	HumanEval pass@1	84.70	84.60	99.9%
	HumanEval+ pass@1	79.50	79.90	100.5%
	MBPP pass@1	71.10	70.10	98.6%
	MBPP+ pass@1	60.60	60.70	100.2%
Vision	MMMU (0-shot)	52.11	50.11	96.2%
	ChartQA (0-shot)	81.36	80.92	99.5%