Mistral Medium 3.5

Generic Info

• Publisher: Mistral AI
• Release Date: 2026
• Architecture: Dense 128B Parameters Transformer with Multimodal Vision Encoder
• Context Window: 256,000 tokens (256k context)
• License: Modified MIT License (Permissive commercial & non-commercial use)
• Key Capabilities: Dual Mode (Instant vs. Reasoning), Image Understanding, First-Class Function Calling, Structured JSON Output

Mistral Medium 3.5 represents Mistral's flagship dense architecture. Built with 128 billion dense parameters, it pairs raw parameter capacity with the agility to switch seamlessly between sub-second conversational replies and deep test-time reasoning. With multimodal input support, a 256k context window, and industry-standard tool calling accuracy, it provides an exceptional backbone for production agent services.

Hello World Guide

Run Mistral Medium 3.5 using vLLM or the official Mistral inference stack.

from vllm import LLM, SamplingParams

model_name = "mistralai/Mistral-Medium-3.5-128B"

# Initialize vLLM with tensor parallelism across GPUs
llm = LLM(
    model=model_name,
    tensor_parallel_size=4,
    max_model_len=32768
)

sampling_params = SamplingParams(
    max_tokens=1024,
    temperature=0.7
)

prompt = "You are a senior system architect. Design an event-driven distributed pipeline for high-frequency telemetry."
outputs = llm.generate([prompt], sampling_params)

for output in outputs:
    print(output.outputs[0].text)

Industry Usage