Mistral's smallest model at 3B parameters. Ultra-fast and cost-efficient for lightweight tasks.
Ministral 3 3B is the smallest member of the Mistral 3 family, released on December 2, 2025. With 3.4 billion language model parameters paired with a 0.4B vision encoder, it delivers robust text and image understanding in a compact package designed for edge deployment. Licensed under Apache 2.0, it is available as both an instruct and reasoning post-trained variant and fits in approximately 8 GB of VRAM in FP8 — making it deployable on consumer hardware with no GPU cluster needed.
Despite its tiny size, Ministral 3 3B delivers competitive reasoning on benchmarks like AIME 2025 (72.1%) and LiveCodeBench (54.8%), outperforming Qwen3 4B in several head-to-head comparisons. It supports agentic capabilities including native function calling and structured JSON output, a 256K context window, multilingual coverage across 10+ languages, and vision input for up to 10 images per prompt. For cost-sensitive, edge, or on-device workloads, it is one of the most capable ultra-compact models available in Q2 2026.
| Type | Rate |
|---|---|
| Input | $0.108 / 1M tokens |
| Output | $0.108 / 1M tokens |
| Release date | 2025-12-02AWS Bedrock model card ↗ | verified |
| Architecture | 3.4B LLM + 0.4B vision encoder = 3.8B total, dense Transformer with GQA (32 query / 8 KV heads)Hugging Face model card ↗ | verified |
| Context window | 256K tokens (262,144)Hugging Face model card ↗ | verified |
| Max output tokens | 8K tokensAWS Bedrock model card ↗ | verified |
| Training cutoff | Not publicly disclosed | unknown |
| License | Apache 2.0Hugging Face — License ↗ | verified |
| Languages | 10+ languages including EN, FR, ES, DE, IT, PT, NL, ZH, JA, KO, ARHugging Face — Languages ↗ | verified |
| Recommended temperature | 0.1Hugging Face — Recommended settings ↗ | verified |
| VRAM requirement (FP8) | ~8 GBHugging Face — Deployability ↗ | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
AIME 2025 Ministral 3 3B (reasoning variant) — leading among sub-4B models, ahead of Qwen3 4B (69.7%) | 72.1%% | Hugging Face — Reasoning benchmarks ↗ | |
AIME 2024 | 77.5%% | Hugging Face — Reasoning benchmarks ↗ | |
GPQA Diamond | 53.4%% | Hugging Face — Reasoning benchmarks ↗ | |
LiveCodeBench Leads among sub-4B models, ahead of Qwen3 4B (51.3%) | 54.8%% | Hugging Face — Reasoning benchmarks ↗ | |
Arena Hard | 30.5%% | Hugging Face — Instruct benchmarks ↗ | |
WildBench | 56.8 | Hugging Face — Instruct benchmarks ↗ | |
MATH Maj@1 | 83.0%% | Hugging Face — Instruct benchmarks ↗ | |
MM MTBench Score normalised to 78.3% for bar display | 7.83 / 10/10 | Hugging Face — Instruct benchmarks ↗ | |
Multilingual MMLU | 65.2%% | Hugging Face — Base benchmarks ↗ | |
MMLU (5-shot) | 70.7%% | Hugging Face — Base benchmarks ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "mistral/ministral-3b",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'Standard chat completions with system prompt and multi-turn conversation support. Stream responses for real-time UX.
curl https://api.therouter.ai/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-d '{
"model": "mistral/ministral-3b",
"messages": [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Classify this sentence: The cat sat on the mat."}
],
"temperature": 0.1,
"max_tokens": 256,
"stream": true
}'Mistral AI released the complete Mistral 3 model family on December 2, 2025, spanning from the ultra-compact Ministral 3 3B to the flagship Mistral Large 3 (675B). The entire family was trained on NVIDIA Hopper GPUs and most members are open-weight under Apache 2.0. The Ministral 3 3B specifically targets edge and on-device deployment with 3.8B total parameters including a vision encoder.
Ministral 3 3B is available in two post-trained variants. The Instruct variant (FP8 weights) is fine-tuned for chat, instruction following, and tool use — it is the recommended default for most production scenarios. The Reasoning variant (BF16 weights) is post-trained for chain-of-thought reasoning, making it better suited for math, coding, and STEM tasks at the cost of slightly higher VRAM requirements.
Yes. In FP8 precision, Ministral 3 3B requires approximately 8 GB of VRAM — fitting a single RTX 4060 (12 GB) or Apple Silicon Mac with at least 8 GB of unified memory. With further quantization (INT4), it can run on even more modest hardware. This makes it one of the most capable models for fully local, offline-capable AI assistants.
Yes. The model has a dedicated 0.4B vision encoder paired with the 3.4B language model, supporting up to 10 images per prompt. Images can be provided as base64-encoded data URLs or remote URLs in the content array. Mistral recommends maintaining an aspect ratio close to 1:1 for optimal performance.
Apache 2.0 — allowing free use, modification, and distribution for both commercial and non-commercial purposes. This is the same open-source license used by the entire Ministral 3 family, making it one of the most permissively licensed capable edge models.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | AWS Bedrock model card ↗ | 2026-05-27 | verified |
| Architecture | Hugging Face model card ↗ | 2026-05-27 | verified |
| Context window | Hugging Face model card ↗ | 2026-05-27 | verified |
| Max output tokens | AWS Bedrock model card ↗ | 2026-05-27 | verified |
| Training cutoff | — | — | unknown |
| License | Hugging Face — License ↗ | 2026-05-27 | verified |
| Languages | Hugging Face — Languages ↗ | 2026-05-27 | verified |
| Recommended temperature | Hugging Face — Recommended settings ↗ | 2026-05-27 | verified |
| VRAM requirement (FP8) | Hugging Face — Deployability ↗ | 2026-05-27 | verified |
| AIME 2025 | Hugging Face — Reasoning benchmarks ↗ | 2026-05-27 | verified |
| AIME 2024 | Hugging Face — Reasoning benchmarks ↗ | 2026-05-27 | verified |
| GPQA Diamond | Hugging Face — Reasoning benchmarks ↗ | 2026-05-27 | verified |
| LiveCodeBench | Hugging Face — Reasoning benchmarks ↗ | 2026-05-27 | verified |
| Arena Hard | Hugging Face — Instruct benchmarks ↗ | 2026-05-27 | verified |
| WildBench | Hugging Face — Instruct benchmarks ↗ | 2026-05-27 | verified |
| MATH Maj@1 | Hugging Face — Instruct benchmarks ↗ | 2026-05-27 | verified |
| MM MTBench | Hugging Face — Instruct benchmarks ↗ | 2026-05-27 | verified |
| Multilingual MMLU | Hugging Face — Base benchmarks ↗ | 2026-05-27 | verified |
| MMLU (5-shot) | Hugging Face — Base benchmarks ↗ | 2026-05-27 | verified |
| Mistral 3 family announced — from 3B to 675B | Mistral AI Blog ↗ | 2026-05-27 | verified |
| What is the difference between the Instruct and Reasoning variants? | Hugging Face — Model variants ↗ | 2026-05-27 | to verify |
| Can Ministral 3 3B run on a laptop? | Hugging Face — Deployability ↗ | 2026-05-27 | to verify |
| Does Ministral 3 3B support vision input? | Hugging Face — Key Features ↗ | 2026-05-27 | to verify |
| What license is Ministral 3 3B released under? | Hugging Face — License ↗ | 2026-05-27 | to verify |