Fast et very cheap
À propos de 360 tokens a second in our tests, en $0.03 per 1M input tokens, for steps that run millions of times.
Tarif public The Ministral 3 3B API: Mistral AI's smallest open-weight model, with vision, for fast, high-volume work, 70% below Mistral's
ministral-3-3bfrom openai import OpenAIclient = OpenAI( base_url="https://api.zurelay.com/v1", api_key="YOUR_ZURELAY_KEY",)stream = client.chat.completions.create( model="ministral-3-3b", messages=[{"role": "user", "content": "Explain quicksort in two sentences."}], stream=True,)for chunk in stream: if chunk.choices: print(chunk.choices[0].delta.content or "", end="")URL de base https://api.zurelay.com/v1. Compatible avec tous les SDK OpenAI. Documentation de l’API de chat
Tarifs
Payez à l’usage avec votre crédit prépayé, sans abonnement. Rechargez dès $10. Les requêtes qui échouent ne sont jamais facturées.
| Tarif | Zurelay | Tarif public Mistral AI | Vos économies |
|---|---|---|---|
Entrée par 1M tokens | $0.03 | 70%de réduction | |
Sortie par 1M tokens | $0.03 | 70%de réduction |
Un tarif unique, quelle que soit la longueur du prompt, up to the full 128K context window. Le streaming, les appels d’outils et les sorties structurées sont sans supplément.
Ajustez les curseurs selon votre utilisation mensuelle.
$50.40 par an
Vue d’ensemble
Ministral 3 3B is the smallest model in Mistral AI's Ministral 3 family, released on December 2, 2025: an open-weight model that reads text et images, made for fast, low-cost work such as classification, extraction, translation et image captions. The Zurelay Ministral 3 3B API serves its instruct version through one OpenAI-compatible endpoint with a 131,072-token context window, at 70% below Mistral's list price.
Mistral AI released the Ministral 3 family on December 2, 2025, alongside Mistral Large 3, in three sizes: 3B, 8B et 14B. Each size comes as a base, an instruct and a reasoning model, all with image understanding and all under the Apache 2.0 license, with the weights on Hugging Face. Ministral 3 3B pairs a 3.4B-parameter language model with a 0.4B-parameter vision encoder, and Mistral says it fits in 8GB of GPU memory at FP8. On Zurelay the model ID is ministral-3-3b, and it coûte $0.03 per 1M input tokens and $0.03 par 1M tokens en sortie.
Mistral built Ministral 3 for edge et local use, et points the 3B model at lightweight, real-time work: image captions, text classification, translation, data extraction and short content. It handles dozens of languages, including English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean and Arabic. Mistral says its Ministral instruct models often write an order of magnitude fewer tokens than comparable models.
On Mistral's model card the instruct model scores 56.8 on WildBench, level with Qwen3-VL-4B-Instruct et ahead of Gemma 3 4B (49.1), et 0.830 on MATH, ahead of Gemma 3 4B (0.759) and behind Qwen3-VL-4B-Instruct (0.900). Qwen3-VL-4B-Instruct also leads it on Arena Hard, 0.438 against 0.305. These are Mistral's numbers, not independent ones.
The instruct model answers directly without reasoning first, so reasoning_effort has no effect. It calls tools, several at once et while streaming, et returns JSON for json_object or json_schema in response_format. Mistral lists a 256K context window on its own API; on Zurelay a request holds up to 131,072 tokens. In our tests it wrote about 360 tokens a second and knew little about events after 2024, as small models tend to, so put the facts it needs in the prompt. You pay $0.03 per 1M input tokens and $0.03 per 1M output tokens against Mistral's list price of $0.10 and $0.10.
Points forts
À propos de 360 tokens a second in our tests, en $0.03 per 1M input tokens, for steps that run millions of times.
A 0.4B-parameter vision encoder reads photos, screenshots et scans next to your text, for captions et quick visual checks.
Appel de fonctions et JSON output that Mistral describes as native, for routing, extraction et small agent steps.
Apache 2.0 weights sur Hugging Face, small enough for 8GB sur GPU memory, so what you build on Zurelay carries over to a local copy later.
Cas d’usage
Sort tickets, tag content et pick the next step in a pipeline, with JSON schema output for the code that follows.
Pull fields from text, forms et screenshots into JSON en a very low price per call.
Quick translation between the languages Mistral lists, including English, French, Spanish, German, Chinese, Japanese et Arabic.
Short descriptions et alt text for photos et screenshots at volume.
Commencer
Pas de liste d’attente ni de nouveau SDK. Si votre code communique déjà avec OpenAI, il peut aussi communiquer avec Zurelay.
Sign up, add credit et create an API key. Set a monthly budget or a rate limit per key if you like.
Change the base URL. Everything else in your code stays the same.
https://api.zurelay.com/v1Send chat completions as usual. Streaming, tool calls et usage reporting work as you expect.
ministral-3-3bCompatible avec les outils que vous utilisez déjà
Comparer
$0.06 / $0.06 par 1M tokens
Pick Ministral 3 14B, the largest Ministral 3 model, for stronger answers: 0.551 against 0.305 on Arena Hard in Mistral's tests, en twice the price per token.
API Ministral 3 14B$0.07 / $0.22 par 1M tokens
Pick Qwen 3.8 Flash for a 1M-token context, video input et a thinking mode when a task needs more than a small model.
API Qwen 3.8 Flash$0.03 / $0.12 par 1M tokens
Pick DeepSeek V4.1 Flash for a 1M-token context et thinking on harder text and image tasks.
API DeepSeek V4.1 FlashFAQ
D’autres questions ? Contactez notre équipe et nous vous répondrons par e-mail.
On Zurelay, ministral-3-3b costs $0.03 per 1M input tokens et $0.03 per 1M output tokens. Mistral's list price is $0.10 input et $0.10 output per 1M, so you save 70%. One rate applies at every prompt length, and failed requests are never billed.
Install the official openai package, set base_url to https://api.zurelay.com/v1 et use your Zurelay API key. Then send a Chat Completions request with model set to "ministral-3-3b". Zurelay also accepts ministral-3-3b-instruct, mistralai/ministral-3-3b et ministral-3b. Mistral recommends a temperature of 0.1 or lower for production work.
On Zurelay a request holds up to 131,072 tokens. Mistral lists a 256K window on its own API. The model reads text et images et writes text.
It reads images: send them in image_url parts, as https links or base64 data URLs, et they count as input tokens. Mistral suggests images close to square, with very wide or thin ones cropped. Zurelay doesn't take PDFs, audio or video for Ministral 3 3B: a request that includes one is turned away with a clear error before anything is billed.
No. Zurelay serves the instruct model, which answers directly, so reasoning_effort has no effect et there are no reasoning tokens à pay for. Mistral also publishes a separate Ministral 3 3B Reasoning model, which Zurelay doesn't serve. For answers with thinking, use Qwen 3.8 Flash or DeepSeek V4.1 Flash.
Yes. Pass tools as you would to any OpenAI model; it can call several at once, also while streaming. A tool forced by name is called, et tool_choice "none" works. Tool call ids must be at least 9 characters long, as the model's own et OpenAI's are. For JSON, send response_format with json_object or json_schema and the answer comes back as JSON.
Both read text et images, hold 131,072 tokens on Zurelay et call tools. 14B pairs a 13.5B-parameter language model with the same 0.4B vision encoder and scores higher in Mistral's tests (0.551 against 0.305 on Arena Hard, 68.5 against 56.8 on WildBench), at twice the price. 3B is faster, about 360 tokens a second against 170 in our tests, and suits simple, high-volume steps.
Its weights are open under the Apache 2.0 license, which allows commercial use et changes. Mistral publishes base, instruct et reasoning versions on Hugging Face and says the 3B fits in 8GB of GPU memory at FP8. The API is the quick way to use it without running your own hardware.
Yes. Requests for ministral-3-3b run on Mistral AI's Ministral 3 3B Instruct, et Zurelay doesn't swap in a different model. Only smart routing, which you turn on yourself, can answer with a close alternative when the model is down, et the x-zurelay-model header names the model that answered. Responses are sampled, so wording varies between runs. Zurelay is independent and is not affiliated with or endorsed by Mistral AI.
Créez une clé en quelques secondes, configurez votre SDK habituel pour utiliser Zurelay et payez chaque requête jusqu’à 90% moins cher, dès le premier token.