magistral_small_2507_en

Model Overview

Model Summary

Magistral is MistralAI's family of reasoning-focused large language models, fine-tuned from Mistral Small 3.1 using supervised fine-tuning (SFT) on reasoning traces and reinforcement learning (RL). Magistral models are designed for complex, multi-step reasoning tasks including mathematics, coding, logic, and scientific analysis.

Key features:

-Architecture: Based on Mistral with explicit head_dim=128 and no sliding window attention

-**Reasoning:**Uses [THINK]/[/THINK] special tokens for chain-of-thought reasoning

-**Context:**128K token context window (recommended up to 40K for optimal performance)

• [Magistral Quickstart Notebook](coming soon..!)
• Magistral API Documentation
• Magistral Model Card
• Magistral Technical Paper
• KerasHub Beginner Guide
• KerasHub Model Publishing Guide

Installation

Keras and KerasHub can be installed with:

pip install -U -q keras-hub
pip install -U -q keras

Jax, TensorFlow, and Torch come preinstalled in Kaggle Notebooks. For instructions on installing them in another environment see the Keras Getting Started page.

Presets

Example Usage

Configure the Backend (Optional but recommended)

import os
# Configure Keras to use the JAX backend 
os.environ["KERAS_BACKEND"] = "jax" 

Load the Model & Generate Text

import keras
import keras_hub

# Load the Magistral Small 2507 model
# This will automatically download the weights from Kaggle
causal_lm = keras_hub.models.MistralCausalLM.from_preset("magistral_small_2507_en")

# Generate text
prompt = "The most important concepts to understand in machine learning are"
output = causal_lm.generate(prompt, max_length=100)

print(output)

Loading just the Backbone

# Load only the feature extractor backbone
backbone = keras_hub.models.MistralBackbone.from_preset("magistral_small_2507_en")

# Pass a batch of tokenized sequences to get hidden states
import numpy as np
dummy_tokens = np.random.randint(0, 1000, size=(2, 64)) 
hidden_states = backbone(dummy_tokens)

print(hidden_states.shape) 

Example Usage with Hugging Face URI

Configure the Backend (Optional but recommended)

import os
# Configure Keras to use the JAX backend 
os.environ["KERAS_BACKEND"] = "jax" 

Load the Model & Generate Text

import keras
import keras_hub

# Load the Magistral Small 2507 model
# This will automatically download the weights from Kaggle
causal_lm = keras_hub.models.MistralCausalLM.from_preset("hf://keras/magistral_small_2507_en")

# Generate text
prompt = "The most important concepts to understand in machine learning are"
output = causal_lm.generate(prompt, max_length=100)

print(output)

Loading just the Backbone

# Load only the feature extractor backbone
backbone = keras_hub.models.MistralBackbone.from_preset("hf://keras/magistral_small_2507_en")

# Pass a batch of tokenized sequences to get hidden states
import numpy as np
dummy_tokens = np.random.randint(0, 1000, size=(2, 64)) 
hidden_states = backbone(dummy_tokens)

print(hidden_states.shape)