Ministral-3-8B-Instruct-2512

The Ministral-3-8B-Instruct-2512 strikes the sweet spot in the Ministral 3 family, offering an optimal balance between computational efficiency and response quality. Its architecture includes a text LLM with 8.4 billion parameters and a vision encoder with 0.4 billion parameters. The model is provided by its developers in an FP8 quantized format. A context window of 256,000 tokens enables processing large volumes of information, and the Apache 2.0 license permits free commercial use.

The Cascade Distillation technology underlying Ministral 3 allows the 8B model to retain a significant portion of the capabilities of its parent model, Mistral Small 3.1 (24B), while reducing parameters by nearly three times. This is achieved through iterative pruning followed by knowledge distillation, which substantially lowers training computational costs without meaningful quality loss. The 410M‑parameter vision encoder works in tandem with an adapter, providing efficient multimodal perception with minimal overhead.

On the Arena Hard benchmark (instruction-following evaluation), the model scores 0.509, which is comparable to Qwen3-VL-8B-Instruct (0.528) and higher than Gemma3-12B-Instruct (0.436). On WildBench (dialogue capabilities), its result of 66.8 surpasses Qwen3-VL-8B-Instruct (66.3). On the MATH Maj@1 benchmark, the model achieves 0.876, demonstrating strong analytical abilities at a relatively compact size.

When using the model, developers are advised to clearly define the environment and use case in the system prompt. Use a temperature below 0.1 for productive environments and minimize the number of tools in agentic scenarios. For visual input, use images with an aspect ratio close to 1:1. The model is well‑suited for local AI assistants and chat interfaces in constrained environments, as well as for image/document description, translation and content generation, and specialized agentic scenarios.

Announce Date: 31.10.2025
Parameters: 9B
Context: 263K
Layers: 34
Attention Type: Full Attention
Developer: Mistral AI
Transformers Version: 5.0.0.dev0
License: Apache 2.0

Use our pre-built public endpoints for free to test inference and explore Ministral-3-8B-Instruct-2512 capabilities. You can obtain an API access token on the token management page after registration and verification.

Rent your own physically dedicated instance with hourly or long-term monthly billing.

We recommend deploying private instances in the following scenarios:

• maximize endpoint performance,
• enable full context for long sequences,
• ensure top-tier security for data processing in an isolated, dedicated environment,
• use custom weights, such as fine-tuned models or LoRA adapters.

Contact our dedicated neural networks support team at nn@immers.cloud or send your request to the sales department at sale@immers.cloud.