Ministral 3 8B review, pricing and limits

Ministral 3 8B suits developers who have outgrown the 3B's reasoning but do not want the 14B's VRAM bill. The dedicated reasoning checkpoint scores 78.7% on AIME 2025, and the whole family runs on a single consumer GPU instead of a hosted API, which is the real practical win here.

Ministral 3 8B is Mistral AI's mid-tier open-weight model, built with a 256,000-token context window and a built-in vision encoder that reads charts and screenshots without a separate model call. It sits between the Ministral 3B and 14B, aimed at reasoning and multilingual tasks that fit on a single consumer GPU.

Provider: Mistral AI · Family: Ministral 3

More about Mistral AI on HokAI

Context window: 256,000 tokens

Input modalities: text, image, tool-calls · Output: text, tool-calls

About Ministral 3 8B

Mistral AI released Ministral 3 8B in December as the mid-tier member of its Ministral 3 edge-model family, above the 3B and below the 14B. It uses a dense Transformer architecture with roughly 8 billion parameters and interleaved sliding-window attention for memory-efficient long-context inference. The weights are open-licensed for commercial use, released as three separate checkpoints: Base, Instruct, and a dedicated Reasoning variant tuned for math and logic. Mistral's own testing found the Reasoning checkpoint competitive with Qwen2.5 14B Instruct, released in September 2024, on math benchmarks despite having roughly half the parameters, while adding vision input Qwen2.5 14B lacks.

The model accepts text and image input through a 410M-parameter Vision Transformer encoder built into the architecture, handling chart reading and OCR without a separate vision call; output is text only, with no audio or video support. Function calling and structured JSON output work through Mistral's OpenAI-compatible tool schema with parallel tool calls, and the model covers 11 languages including French, German, Japanese, and Chinese. In BF16 the model needs 24GB of VRAM; the FP8 checkpoint drops that to 12GB, small enough for a single RTX 3080 or RTX 4070, and a Q4 GGUF build runs under 6GB on hardware like an Apple M2 Pro.

Safety alignment follows standard SFT plus RLHF; Mistral has not published a separate red-team report for the 8B, and its default refusals are lighter than Claude or GPT-4o Mini, so production deployments typically add their own content filter. API traffic through la Plateforme is not used for training, and Mistral AI is a GDPR-compliant European company routing that traffic through EU infrastructure, with enterprise data agreements available through its sales team.

Pricing

$0.15 per 1M input tokens and $0.15 per 1M output tokens via la Plateforme, with no published cached-input discount. Self-hosted deployments under an open license pay only their own compute costs; Mistral does not charge a per-token fee for downloaded weights.

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• 410M-Parameter Vision Encoder: A built-in Vision Transformer reads images, charts, and screenshots directly, so visual QA and OCR do not need a separate model call.
• 66.8% GPQA Diamond (Reasoning Variant): The dedicated reasoning checkpoint scores 66.8% on GPQA Diamond and 61.6% on LiveCodeBench, ahead of many instruct-only models at this parameter count.
• Runs in 12GB of VRAM: The FP8 checkpoint fits on a single RTX 3080 or RTX 4070, so reasoning-grade inference does not require a data-center GPU.
• 256,000-Token Context: Handles long documents and multi-turn chats via interleaved sliding-window attention, twice the window of comparably sized dense models.
• Three Variants, One License: Base, Instruct, and Reasoning checkpoints are all released under an open license on Hugging Face, so teams can pick the right checkpoint per task without extra fees.

Pros

• The reasoning checkpoint's 66.8% GPQA Diamond score is a genuine standout for an 8B model, not just a marketing number.
• Cheaper per token than comparable hosted 8B alternatives while requiring no data-center hardware to self-host.
• Native vision input means teams do not need to bolt on a separate OCR or vision model for document-heavy agents.

Cons

• Math-heavy benchmark scores apply to the dedicated Reasoning checkpoint only; the Instruct model needs more prompting to match them.
• BF16 needs 24GB of VRAM; FP8 (12GB) or Q4 (under 6GB) is required for smaller cards.
• No native audio input, so voice applications need a separate speech-to-text step before calling the model.

Benchmarks

• MMLU-Pro: 70.6% vendor-reported · 02 Dec 2025 — A harder version of the 57-subject knowledge exam, % correct.
• AIME 2025: 78.7% vendor-reported · 02 Dec 2025 — Competition-level maths problems from the 2025 exam, % solved.
• LiveBench: 61.6% vendor-reported · 02 Dec 2025 — A rolling set of fresh questions that cannot have been in training data, % correct.
• GPQA Diamond: 66.8% vendor-reported · 02 Dec 2025 — PhD-level science questions that are hard to search for, % correct.
• AA blended price: $0.15/M cited: Artificial Analysis — Price per 1M tokens at a 3:1 input to output blend, as listed by Artificial Analysis.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit Ministral 3 8B Official Page