Magistral Mistral AI Released 2026-06 reasoningthinkingsmall

Magistral

Magistral is a small, efficient reasoning model with 24B parameters. From Mistral AI. Best for hardware that cannot run larger reasoning models.

Best for small efficient reasoning model
Sizes 24B
Context 32K
License MRL License (open-weights)
Min VRAM (default size, Q4) 15 GB
Rec VRAM 24 GB

What is Magistral?

Magistral is Mistral AI's open-weights language model. Released in 2026-06, it ships in 1 size (24B) and is licensed MRL License (open-weights). The most popular use case is small efficient reasoning model.

VRAM and hardware

The 24B size fits in 15 GB of VRAM.

Size Min VRAM (Q4_K_M) Recommended VRAM Tokens/sec on 3090
24B 15 GB 23 GB ~19 tok/s

How to run Magistral locally

Option 1: Ollama (simplest)

ollama pull magistral
ollama run magistral

Option 2: Mullama (production)

mullama pull magistral
mullama run magistral

Option 3: llama.cpp (CLI)

# Download a GGUF from Hugging Face (search "magistral gguf")
./llama-cli -m magistral.Q4_K_M.gguf -p "Hello, AI!"

Option 4: Python with Mullama or llama-cpp-python

from mullama import Model, Context
model = Model.load("magistral.Q4_K_M.gguf", n_gpu_layers=99)
ctx = Context(model, n_ctx=4096)
print(ctx.generate("Hello, AI!", 256))

What you can build with Magistral

  • reasoning
  • math
  • small chat

Hardware it fits on

  • Apple Silicon (8GB): 24B in MLX
  • Apple Silicon (16GB): 24B in MLX
  • Apple Silicon (32GB): 24B in MLX or GGUF
  • Apple Silicon (64GB+): 24B in MLX
  • RTX 3090 (24GB): 24B at Q4_K_M
  • RTX 4090 (24GB): same as 3090, ~30% faster
  • RTX 5090 (32GB): 24B at Q8_0
  • 2× RTX 6000 Ada (96GB total): 24B at Q4_K_M

Related models in the Magistral family

This is the only Magistral-family model in the catalog right now.

Other models you might consider

  • Llama 3.1 (Meta) — 8B, general-purpose chat, agentic workflows, code
  • Llama 3.2 (Meta) — 1B, low-end hardware, mobile, edge deployment
  • Llama 3.2 Vision (Meta) — 11B, vision + chat on desktop GPUs
  • Llama 3.3 (Meta) — 70B, best-in-class 70B for desktops
  • Llama 4 (Meta) — Scout 17B (109B MoE), vision + long context (10M tokens)
  • Qwen 2.5 (Alibaba) — 0.5B, strongest 7B-32B on consumer hardware