Methods to entry Ministeral 3 through API
TL;DR
minister 3 is a household of fashions optimized for open-weight inference, obtainable in each 3B and 14B variants. These fashions assist multimodal inference, calling native capabilities and instruments, and an enormous 256K token context window, all launched underneath the Apache 2.0 license.
To run Ministral 3 immediately on Clarifai, playground You need to use it for interactive testing or combine it into your purposes via Clarifai’s OpenAI-compatible API.
This information describes the Ministeral 3 structure, methods to entry it via Clarifai, and the way to decide on the correct variant to your manufacturing workloads.
introduction
Trendy AI purposes more and more depend on fashions that may purpose reliably, preserve long-term context, and combine cleanly into present instruments and APIs. Whereas closed-source fashions have traditionally led these capabilities, open-source alternate options are quickly closing the hole.
Among the many open fashions obtainable worldwide, Ministeral 3 ranks among the many prime together with the DeepSeek and GPT OSS households. Quite than concentrating on leaderboard efficiency in benchmarks, Ministral prioritizes efficiency that issues in manufacturing, reminiscent of producing structured output, processing massive paperwork, and making operate calls inside a stay system.
This makes Ministeral 3 appropriate for the calls for of real-world enterprise purposes, as organizations more and more undertake an open weight mannequin for transparency, deployment flexibility, and the power to run throughout a wide range of infrastructure configurations, from cloud platforms to on-premises techniques.
ministral 3 structure
Ministral 3 is a household of dense, edge-optimized, multimodal fashions designed for environment friendly inference, long-context processing, and native or personal deployment. This household presently consists of 3B and 14B parameter fashions, obtainable in base, instruction, and inference variants, respectively.
Minister 3 14B
The biggest mannequin within the Ministral household is a dense inference-posttraining structure optimized for math, coding, STEM, and different multi-step inference duties. It combines a ~13.5B parameter language mannequin with a ~0.4B parameter imaginative and prescient encoder to allow native textual content and picture understanding. The 14B inference variant achieves 85% accuracy on AIME ’25, delivering state-of-the-art efficiency inside its weight class whereas being deployable on practical {hardware}. It helps context home windows of as much as 256,000 tokens and is appropriate for lengthy paperwork and sophisticated inference workflows.
Ministral 3 3B
The 3B mannequin is a compact, inferential, post-training variant designed for extremely environment friendly deployment. It combines a language mannequin with roughly 3.4B parameters and a imaginative and prescient encoder with roughly 0.4B parameters (complete roughly 4B parameters) to supply multimodal performance. Much like the 14B mannequin, it helps a context size of 256k tokens, enabling lengthy context inference and doc evaluation on constrained {hardware}.
Important technical options
Multimodal operate: All Ministral 3 fashions use a hybrid language and imaginative and prescient structure that may course of textual content and pictures concurrently for duties reminiscent of doc understanding and visible reasoning.
Lengthy context reasoning: The inference variant helps as much as 256,000 tokens and permits long-running conversations, large-scale doc ingestion, and multi-step evaluation workflows.
Environment friendly inference: The mannequin is optimized for edge and personal deployments. The 14B mannequin runs on BF16 with roughly 32 GB VRAM, whereas the 3B mannequin runs on BF16 with roughly 16 GB VRAM, and the quantized model requires considerably much less reminiscence.
Agent workflow: Ministral 3 is designed to work effectively with structured output, operate calls, and using instruments, making it appropriate for automation and agent-based techniques.
license: All variants of Ministral 3 are launched underneath the Apache 2.0 license, permitting limitless business use, tweaking, and customization.
Pre-training benchmark efficiency
The Ministral 3 14B demonstrates sturdy reasoning means and multilingual efficiency in comparison with equally sized open fashions, whereas sustaining aggressive outcomes on normal data duties. It’s significantly good at reasoning-oriented benchmarks, demonstrating strong factual copy and multilingual understanding.
benchmark
Minister 3 14B
Gemma 3 12B base
Qwen3 14B base
Precautions
math cot
67.6
48.7
62.0
A robust lead on structural reasoning
MMLU Redux
82.0
76.6
83.7
aggressive normal data
Trivia QA
74.9
78.8
70.3
Reproducing sure information
Multilingual MMLU
74.2
69.0
75.4
Highly effective multilingual efficiency
Entry Ministral 3 through Clarifai
Stipulations
Earlier than you’ll be able to run Ministral 3 utilizing the Clarifai API, you will need to full some primary setup steps.
Make clear account: create Make clear An account to entry hosted AI fashions and APIs.
Private Entry Token (PAT): All API requests require a private entry token. You’ll be able to generate or copy from . Settings > Secret A piece of the Clarifai dashboard.
For added SDKs and setup steerage, see under. Clarifai Fast Begin Doc.
API utilization
The instance under makes use of the biggest mannequin within the Ministeral 3 household, Ministral-3-14B-Reasoning-2512. Optimized for multi-step inference, mathematical drawback fixing, and long-context workloads, it is preferrred for lengthy doc use circumstances and agent purposes. Right here we present you methods to make your first API name to your mannequin utilizing numerous strategies.
Python (OpenAI suitable)
Python (Clarifai SDK)
It’s also possible to use the Clarifai Python SDK for extra management over technology settings and inference. This is methods to use the SDK to make predictions and produce streaming output.
Node.js (Clarifai SDK)
This is methods to carry out inference utilizing the Node.js SDK.
playground
of Make clear Playground Rapidly experiment with prompts, structured output, inference workflows, and performance calls with out writing any code.
Please go to. playground Then select one of many following:
Ministry-3-3B-Inference-2512
Ministry-3-14B-Inference-2512
Functions and use circumstances
Ministral 3 is designed for groups constructing clever techniques that require highly effective inference, long-term contextual understanding, and dependable, structured output. Superior efficiency throughout agent, technical, multimodal, and business-critical workflows.
agent utility
Ministral 3 is right for AI brokers that have to plan, purpose, and act throughout a number of steps. The structured JSON output can be utilized to tune instruments and APIs, making it dependable for automation pipelines the place consistency is essential.
lengthy context
Ministral 3 can analyze massive paperwork utilizing an expanded 256K token context, making it efficient for summarizing lengthy technical paperwork, extracting info, and answering questions.
multimodal inference
Ministral 3 helps multimodal reasoning, permitting purposes to mix textual and visible enter in a single workflow. That is helpful for image-based queries, doc understanding, or assistants that have to purpose about combined enter.
conclusion
Ministral 3 supplies an inference-optimized open-weight mannequin that can be utilized in manufacturing. That includes a 256K token context window, multimodal enter, native instrument calls, and OpenAI-compatible API entry via Clarifai, it supplies a sensible basis for constructing superior AI techniques.
The 3B variant is right for low-latency, cost-sensitive deployments, whereas the 14B variant helps deeper analytical workflows. Mixed with the Apache 2.0 license, Ministeral 3 supplies flexibility, efficiency, and long-term management to your workforce.
First, take a look at the next fashions. Make clear Playground Or combine immediately into your utility utilizing the API.
