Mistral Small 24B Instruct 2501 (FP32) — 19.4 GBon NVIDIA RTX 4000 SFF

FP32 NVIDIA RTX 4000 SFF

Overview

Mistral Small 24B Instruct 2501 is a 23.57B parameter dense language model by Mistral AI, with code, multilingual, tool-calls capabilities. It supports a context window of up to 32,768 tokens.

Mistral Small 24B Instruct 2501 is a 23.57-billion-parameter dense transformer from Mistral AI, optimized for instruction following, code generation, and multilingual conversation. It occupies a mid-range parameter class that offers strong performance relative to its size, competing with larger 30B models on many benchmarks. The model supports tool calling and 10 languages including English, French, Chinese, and Japanese. With a 32K context window and flash attention, it fits on a single consumer GPU at Q4 quantization for efficient self-hosted inference.

At FP32 quantization (full-precision quality tier), the model weighs 87.82 GB. This exceeds the 20 GB of VRAM on NVIDIA RTX 4000 SFF. Inference is still possible via CPU offload or memory-mapped loading from disk, but expect significantly reduced performance.

Hardware Requirements

Model size	87.82 GB
VRAM available	20 GB
VRAM used	19.4 GB
Min RAM required	70.3 GB
GPU layers	8 / 40
Context size	3,099
Backend	cuda13
Flash attention	Yes

Performance Notes

Reduced performance

32 of 40 layers run on CPU. Your system needs at least 70.3 GB RAM. If system RAM is insufficient, data will be read from disk on demand, with significant performance impact.

Shorter conversations

Context size has been reduced due to memory constraints. This may affect the model's ability to handle long conversations.

Deploy

Server Local

Prerequisites

Ensure your GPU nodes are prepared with the NVIDIA container toolkit:

ansible-playbook prositronic.infra.nvidia_container_toolkit

Command

helmfile --state-values-file <(curl -s https://www.prositronic.eu/values/mistral-small-24b-instruct-2501/fp32/nvidia-rtx4000sff.yaml) apply

Generated values.yaml

/values/mistral-small-24b-instruct-2501/fp32/nvidia-rtx4000sff.yaml

Loading values…

Install llama.cpp

Install llama.cpp from the official build instructions for your platform and backend.

Download Model

curl -L -o mistral-small-24b-instruct-2501.gguf "https://huggingface.co/bartowski/Mistral-Small-24B-Instruct-2501-GGUF/resolve/main/Mistral-Small-24B-Instruct-2501-f32/Mistral-Small-24B-Instruct-2501-f32-00001-of-00003.gguf"

Start Server

llama-server \
  -m mistral-small-24b-instruct-2501.gguf \
  --n-gpu-layers 8 \
  --ctx-size 3099 \
  --flash-attn

Verify

curl http://localhost:8080/health

Explore

• Configure a deployment
• Browse models
• Browse hardware

Other quantizations

• Mistral Small 24B Instruct 2501 (FP16)
• Mistral Small 24B Instruct 2501 (Q8_0)
• Mistral Small 24B Instruct 2501 (Q5_K_M)

Other hardware

• Mistral Small 24B Instruct 2501 on CPU Only
• Mistral Small 24B Instruct 2501 on NVIDIA A100 80GB
• Mistral Small 24B Instruct 2501 on NVIDIA H100

Frequently Asked Questions

How much VRAM does Mistral Small 24B Instruct 2501 (FP32) need?

The FP32 quantization of Mistral Small 24B Instruct 2501 requires 87.82 GB. 8 of 40 layers fit in the 20 GB of VRAM on NVIDIA RTX 4000 SFF; remaining layers are offloaded to CPU.

Can I run Mistral Small 24B Instruct 2501 on NVIDIA RTX 4000 SFF?

Yes, with reduced performance. NVIDIA RTX 4000 SFF can run Mistral Small 24B Instruct 2501 (FP32), but only 8 of 40 layers fit in VRAM. The rest are offloaded to CPU.

What is quantization?

Quantization reduces a model's numerical precision from its original floating-point format to a more compact representation. This shrinks the file size and VRAM footprint, making it possible to run large models on consumer hardware. The trade-off is a small reduction in output quality. FP32 compresses Mistral Small 24B Instruct 2501 from its original size down to 87.82 GB.

What quantization should I choose for Mistral Small 24B Instruct 2501?

FP32 is a full-precision quantization. Higher-quality quants (Q8, Q6) preserve more model accuracy but need more VRAM. Lower quants (Q4, Q3, Q2) reduce VRAM usage at the cost of some quality. Choose based on your available hardware and quality requirements.

What is flash attention and why is it enabled?

Flash attention is a memory-efficient algorithm that speeds up the attention mechanism in transformer models. It reduces VRAM usage during inference by avoiding the materialisation of the full attention matrix. For Mistral Small 24B Instruct 2501 on NVIDIA RTX 4000 SFF, flash attention is enabled to maximise context length and throughput within the available 20 GB of VRAM.

Why are some layers offloaded to CPU?

NVIDIA RTX 4000 SFF has 20 GB of VRAM, but Mistral Small 24B Instruct 2501 (FP32) requires approximately 87.82 GB. Only 8 of 40 layers fit in VRAM; the remaining layers run on CPU, which is slower but still functional.

How do I run Mistral Small 24B Instruct 2501 (FP32) with Ollama?

Run ollama run mistral-small:24b-instruct-2501-fp32 to start Mistral Small 24B Instruct 2501 (FP32). Ollama handles downloading the model weights automatically on first run.

Last updated: March 12, 2026