Mistral's efficient 14B parameter model with vision support. Good balance of capability and speed for everyday tasks.
Ministral 14B is the largest dense model in Mistral's Ministral 3 family, released December 2, 2025 as part of the broader Mistral 3 launch. It packs 14 billion parameters into a single dense architecture with built-in vision (image understanding) and supports a 256K-token context window on the Mistral API. All Ministral 3 variants ship under the Apache 2.0 license, making them fully open-weight for both research and commercial use.
Mistral positions Ministral 14B as delivering performance comparable to the larger Mistral Small 3.2 (24B) while being optimized for edge and local deployment. Through TheRouter, it serves as a strong mid-tier option: more capable than the 3B/8B siblings for complex reasoning and tool use, yet significantly cheaper and faster than flagship models — ideal for teams that need a solid general-purpose model at modest cost.
| Type | Rate |
|---|---|
| Input | $0.216 / 1M tokens |
| Output | $0.216 / 1M tokens |
| Release date | 2025-12-02mistral.ai ↗ | verified |
| Architecture | Dense transformer, 14B parametersmistral.ai ↗ | verified |
| Training cutoff | Not publicly disclosed | unknown |
| License | Apache 2.0mistral.ai ↗ | verified |
| Model variants | Base, Instruct, and Reasoningmistral.ai ↗ | verified |
| Quantization support | Q4_K_M, FP8, BF16NVIDIA Developer Blog ↗ | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
Performance vs Mistral Small 3.2 (24B) Mistral states Ministral 14B delivers performance comparable to the larger Mistral Small 3.2 24B | Comparable | mistral.ai ↗ |
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-14b",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "mistral/ministral-14b",
"messages": [{"role": "user", "content": "Explain cascade distillation in two paragraphs."}]
}'Mistral ships the Mistral 3 family: three Ministral dense models (14B/8B/3B) and a 675B MoE flagship, all Apache 2.0 licensed with vision support. The 14B sits at the top of the Ministral lineup, targeting edge-to-cloud deployment with performance rivaling 24B-class models.
NVIDIA's developer blog details optimized deployment of all Mistral 3 models across GB200 NVL72 to consumer RTX GPUs, confirming Q4_K_M/FP8/BF16 quantization options for Ministral 14B and compatibility with vLLM, Llama.cpp, and Ollama.
Ministral 14B is part of the Ministral 3 family optimized for edge and local deployment, while Mistral Small (including Small 3.2 at 24B and Small 4) targets cloud-hosted production workloads. Mistral states the 14B delivers comparable performance to Small 3.2 despite having 40% fewer parameters, making it the better choice when hardware resources are constrained.
Yes. All Ministral 3 models are released under Apache 2.0 with open weights. The 14B Instruct variant runs on consumer GPUs with as little as 16 GB VRAM using Q4_K_M quantization via Ollama, Llama.cpp, or vLLM. NVIDIA confirms compatibility with RTX and Jetson hardware.
Yes. All Ministral 3 models — including 14B — ship with built-in image understanding capabilities. Through TheRouter's API, you can send images as part of multi-modal chat completions using the standard OpenAI-compatible vision format.
Mistral states the Mistral 3 family supports 100+ languages. European languages — particularly French, German, Spanish, and Italian — are among its strongest, consistent with Mistral's French AI heritage. Chinese and other Asian languages are supported but may not match dedicated models from Alibaba or Baidu.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | mistral.ai ↗ | 2026-05-28 | verified |
| Architecture | mistral.ai ↗ | 2026-05-28 | verified |
| Training cutoff | — | — | unknown |
| License | mistral.ai ↗ | 2026-05-28 | verified |
| Model variants | mistral.ai ↗ | 2026-05-28 | verified |
| Quantization support | NVIDIA Developer Blog ↗ | 2026-05-28 | verified |
| Performance vs Mistral Small 3.2 (24B) | mistral.ai ↗ | 2026-05-28 | to verify |
| Mistral 3 family launches with Ministral 14B, 8B, 3B and Large 3 | mistral.ai ↗ | 2026-05-28 | verified |
| NVIDIA publishes Mistral 3 acceleration and deployment guide | NVIDIA Developer Blog ↗ | 2026-05-28 | verified |
| What is the difference between Ministral 14B and Mistral Small? | mistral.ai ↗ | 2026-05-28 | to verify |
| Can I run Ministral 14B locally? | NVIDIA Developer Blog ↗ | 2026-05-28 | to verify |
| Does Ministral 14B support image input (vision)? | mistral.ai ↗ | 2026-05-28 | to verify |
| What languages does Ministral 14B support? | mistral.ai ↗ | 2026-05-28 | to verify |