Ministral 3 14B suits teams that want strong open-weight reasoning without renting a large proprietary model: the base model's 74.2% MMLU-Pro score is the highest in its family, and Apache licensing lets you self-host it freely. Skip it for real-time voice work, since there is no native audio input.
Ministral 3 14B is Mistral AI's open-weight multimodal model; its dedicated reasoning variant scores 85% on the AIME math benchmark, the strongest published result in Mistral's small-model family. It reads text and images through an embedded vision encoder and outputs text only.
Where it sits
- $0.2/M$ per 1M tokensBlended price (3:1)Lower is better#10 / 64peer median $1.70/Mvendor price, checked by HokAI
- 92 tok/stokens/sOutput speedHigher is better#18 / 39peer median 90 tok/scited: Artificial Analysis
- --% solvedSWE-bench VerifiedHigher is better-- / 28peer median 78.3%per source, see benchmark scores
- 71.2%% correctGPQA DiamondHigher is better#38 / 44peer median 88.3%per source, see benchmark scores
Cheaper than 86% of the 64 GA models with a published price, in the bottom third on GPQA Diamond (rank 38 of 44), and one of 65 whose vendor states it does not train on customer data.
Ranks are against GA models on HokAI that publish the same figure; ties share a rank.
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 14B
Ministral 3 14B is the largest model in Mistral AI's Ministral 3 family, released December 4, 2025 under the Apache 2.0 license. Paris-based Mistral AI ships it in three variants on Hugging Face: a base model, an instruct model, and a dedicated reasoning model, all distilled from the same recipe.
The dense Transformer architecture combines a 13.5 billion parameter language decoder with a 410 million parameter Vision Transformer encoder, for roughly 14 billion parameters total. Grouped Query Attention across 40 layers with a 5,120 hidden dimension keeps memory use in check without the added complexity of a Mixture-of-Experts design.
On AIME, the reasoning variant scores several points higher than Qwen2.5 14B Instruct's published result, and the base model's MMLU-Pro score is the highest yet published in the Ministral 3 family. Mistral's own Hugging Face model card additionally lists a GPQA Diamond science score for the reasoning variant, a benchmark Qwen2.5 14B Instruct does not publish.
The 256,000 token context window uses the same effective size as Mistral Large 3. Inputs are text and image via the embedded vision encoder; outputs are text only, with function calling and structured JSON output support for agentic workflows. Audio and video are not supported.
In BF16 the model needs 32GB of VRAM; FP8 brings that down to 24GB for a single RTX 4090 or RTX 3090, and Q4 quantization drops it under 10GB for smaller GPUs. Weights are on Hugging Face and the model is also available through NVIDIA Build.
Alignment follows SFT plus RLHF on the instruct variant, with lighter default refusals than many proprietary chat models; Mistral expects deployers to add their own content policies for consumer-facing use. API inputs sent through la Plateforme are not used for training, and self-hosted deployments have no external data exposure.
Pricing
Ministral 3 14B costs $0.20 per 1 million input tokens and $0.20 per 1 million output tokens via Mistral's la Plateforme API, with no published discount for cached or repeated prompts. Self-hosted deployments pay only their own compute cost.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.0060 | $0.0002 | $0.0062 |
| Support reply | $0.0004 | $0.0001 | $0.0005 |
| One coding agent run | $0.040 | $0.0040 | $0.044 |
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
- 85% on AIME 2025 (Reasoning Variant): The dedicated reasoning model ID scores 85% on the AIME 2025 math competition benchmark, Mistral's highest published result in the Ministral 3 family.
- 71.2% on GPQA Diamond (Reasoning Variant): The same reasoning model ID scores 71.2% on GPQA Diamond, a graduate-level science question set, per Mistral's own Hugging Face model card.
- 74.2% MMLU-Pro (Base Model): The base model scores 74.2% on MMLU-Pro, the highest result in the Ministral 3 family according to independent evaluation on LayerLens/Atlas.
- 256K Context Window: Supports a 256,000 token context window using grouped-query attention, matching Mistral Large 3's window at a fraction of the parameter count.
- 24GB FP8 Self-Hosting: Runs in FP8 on a single RTX 4090 or 3090 with 24GB of VRAM, the largest Ministral 3 variant that fits on one consumer flagship card.
- Apache 2.0 with Three Variants: Base, instruct, and reasoning variants are released under Apache 2.0, Mistral's most permissive license, enabling unrestricted commercial self-hosting.
Pros
- Strong reasoning benchmark results at the 14B parameter class, competitive with much larger proprietary models on math and science tasks.
- Runs on a single consumer flagship GPU once quantized to FP8, unlike larger open-weight reasoning models that need multi-GPU setups.
- Includes native vision input in the same model checkpoint, avoiding a separate OCR or captioning pipeline some competitors require.
Cons
- The full-precision BF16 weights need more VRAM than a single consumer flagship GPU provides; a quantized build is required to self-host on one card.
- The reasoning benchmark results apply to a separate model ID from the standard instruct variant, so teams must route math and logic tasks to the right endpoint.
- No native audio input, so voice applications need a separate speech-to-text step before calling the model.
Benchmarks
- MMLU-Pro: 74.2% independent · 06 Sep 2026 — A harder version of the 57-subject knowledge exam, % correct.
- AIME 2025: 85% vendor-reported · 06 Sep 2026 — Competition-level maths problems from the 2025 exam, % solved.
- GPQA Diamond: 71.2% vendor-reported · 06 Sep 2026 — PhD-level science questions that are hard to search for, % correct.
- AA Intelligence Index: 6 cited: Artificial Analysis · 06 Sep 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
- AA blended price: $0.2/M cited: Artificial Analysis · 06 Sep 2026 — Price per 1M tokens at a 3:1 input to output blend, as listed by Artificial Analysis.
- Output speed: 92 tok/s cited: Artificial Analysis · 06 Sep 2026 — Median tokens written per second as measured 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.
Frequently Asked Questions
How much does Ministral 3 14B cost in 2026?
Mistral bills Ministral 3 14B at a flat $0.20 rate per million tokens, whether input or output, through la Plateforme. There is no cache discount the way some rivals offer. Running the open weights yourself sidesteps per-token billing entirely, leaving just your own hardware or cloud compute bill.
Does Ministral 3 14B have a free plan?
Ministral 3 14B has no free tier or trial credits on Mistral's own API. The only no-cost path is self-hosting the openly licensed weights on your own hardware, which avoids per-token charges but requires a GPU capable of running the model.
What are Ministral 3 14B's closest competitors?
Open-weight alternatives at a similar parameter count include Qwen2.5 14B Instruct, Llama 3.1 8B, and Google's Gemma 3 12B. Each trades some combination of context length, licensing terms, or reasoning strength for a different balance than Ministral 3 14B offers.
What separates Ministral 3 14B from Qwen2.5 14B?
Both are small open-weight models, but Ministral 3 14B adds a native vision encoder and a dedicated reasoning variant that publishes a GPQA Diamond science score of 71.2%, a benchmark Qwen2.5 14B Instruct does not report. Qwen2.5 14B Instruct's published context window is 128,000 tokens, about half of Ministral 3 14B's window. Ministral 3 14B is released under Mistral's Apache license terms, while Qwen2.5 uses Alibaba's own community license.
How do you get started with Ministral 3 14B?
Start with an API key from Mistral's la Plateforme console and call the ministral-3-14b-reasoning-2512 model ID for math and logic tasks, or the instruct ID for chat and summarization. Self-hosting requires vLLM or a similar inference server, and enough VRAM to run FP8 weights, or a Q4 quantized build for smaller GPUs.