Mistral · Llm model

Mistral Large 3

Mistral Large 3 is Mistral AI's flagship general-purpose model

Generate on Coinis Read the docs

One API key. Every model.

per 1m tokens
Example
AI text
Commercial use included Verified May 26, 2026 Outputs are yours No training on your data
Endpoints

Start building with Mistral Large 3.

One model. Three ways to call it. Same key, same bill.

IN

Llm

One call. Same key. Same bill.

$0.85 / 1m tokens

OUT

Llm

One call. Same key. Same bill.

$2.55 / 1m tokens

Capabilities

What it does best.

Mistral AI capabilities at a glance

  • Sparse MoE architecture: 41B active parameters out of 675B total, enabling flagship reasoning without activating the full parameter count on every token[^1]
  • 262,144-token (262K) context window, with max output matching total context at 262K tokens, covering large documents and long agentic sessions in one request[^1]
  • Throughput approximately 42 tokens/second, time-to-first-token approximately 0.64 seconds on the Mistral provider[^2]
  • Multimodal inputs: accepts both text and image inputs via image_url, produces text output[^1]
  • Multilingual: French, German, Spanish, Italian, Portuguese, Arabic, Hindi, Russian, Chinese, Japanese, Korean, and more[^3]
  • GPQA Diamond (graduate-level scientific reasoning): 68.0%[^2]
  • SciCode (Python scientific computing): 36.2%[^2]
  • IFBench (instruction-following): 36.2%[^2]
  • Tool/function calling with an observed error rate of approximately 4.09%[^2]
  • Structured JSON output (response_format) with an observed error rate of approximately 0.45%[^2]
  • Streaming, seed-based deterministic sampling, and full OpenAI-compatible request shape[^1]
  • Apache 2.0 open weights. auditable, fine-tunable, and self-hostable under open-source terms[^3]
  • Prompt training disabled: user prompts are not used to train the model[^2]
API

Call Mistral Large 3 in three lines.

One key. One base URL. Same SDK shape you already use. 

# 1. set your key
export COINIS_API_KEY="sk_live_..."

# 2. call the model
curl https://api.app.coinis.com/v1/llm/generate \
  -H "Authorization: Bearer $COINIS_API_KEY" \
  -d '{"prompt":"neon city, rain, tracking shot"}'
import { Coinis } from "@coinis/sdk";
const coinis = new Coinis(process.env.COINIS_API_KEY);

const job = await coinis.llm.generate({
  model: "models/mistral/large-3",
  prompt: "neon city, rain, tracking shot",
});
from coinis import Coinis
coinis = Coinis(os.environ["COINIS_API_KEY"])

job = coinis.llm.generate(
    model="models/mistral/large-3",
    prompt="neon city, rain, tracking shot",
)
Response 
{
  "id": "gen_8fa2c1",
  "status": "succeeded",
  "model": "models/mistral/large-3",
  "output": {
    "image_url": 
                "https://cdn.coinis.com/gen_8fa2c1.mp4"
              
              ,
    "format": "mp4"
  },
  "tokens_used": 10
}

Already on another provider's SDK? Change the host. Keep the call.

Pricing

Token pricing. No surprises.

One wallet across every model. No API accounts to juggle.

Mistral Large 3 · IN
8.5 tokens
per 1m tokens · $0.85
Frontier LLM
$0.85 / 1m tokens
One key. Every model. One invoice. 1 token = $0.10
1 1m tokens ≈ 9 tokens ($0.85)
Start free. 15 tokens a week.

No credit card.

Why pay through Coinis
  • One wallet for every model. No API keys. No separate bills.
  • Generate ads. Launch to Meta. Track in one place.
  • On-brand output from your Brand Profile.

1 token = $0.10 pay-as-you-go. Less on a plan.

Standard vs Fast

Pick the run for the job.

IN

Final renders, studios
Resolution —
Price $0.85 / 1m tokens

OUT

Rapid tests, high volume
Resolution —
Price $2.55 / 1m tokens
Use cases

Two buyers. One model.

For builders

Resell every model. One key. One bill.

Unified API across video, image, audio, and LLM.

Generate 500 variants overnight.

Async queue plus webhooks. Batch at scale.

White-label the output.

Ship it under your brand. Outputs are yours.

For creatives

Ship a Reel before lunch.

Prompt to platform-native clip in minutes.

Same product. Ten formats.

One generation, every aspect ratio.

Commercial UGC without a creator.

Authentic selfie-style ads, on brand.

Agentic workflows Mistral AI's own FAQ recommends Mistral Large for most tasks. The 262K context window and reliable tool-calling fit multi-step task pipelines that need consistent JSON tool responses across many sequential calls.[^3]

Long-document and RAG applications Feed entire contracts, research papers, or codebases into a single 262K-token context. Retrieval-augmented generation over large corpora becomes simpler when chunking is less aggressive.[^1]

Multilingual production apps with EU data requirements Mistral AI is Europe's leading frontier lab. Teams building customer-facing apps in French, German, Spanish, Arabic, or a dozen other languages get flagship-tier quality with a governance story that US-only labs cannot match.[^3]

Multimodal document and image Q&A Pass images alongside text prompts for visual question answering, document comprehension, and screenshot-based analysis. No separate vision model required.[^1]

Cost-sensitive migration off older flagship models Mistral Large 3 is priced at $0.85/1M input tokens and $2.55/1M output tokens on Coinis. Mistral Large 2 on OpenRouter runs $2.00/1M input and $6.00/1M output[^4], roughly 4 times more per token for the older generation. Teams migrating from GPT-4-class pricing see a comparable reduction.

Renders in seconds. Set a seed. Get the same frame back.

Outputs are yours. Sell them.

Safe for paid ads.

Your prompts are never used for training.

FAQ

Mistral Large 3 FAQs

How much does Mistral Large 3 cost on Coinis vs. calling the Mistral API directly?

On Coinis, Mistral Large 3 is priced at $0.85 per 1M input tokens (variant mistral-large-3-in) and $2.55 per 1M output tokens (variant mistral-large-3-out). Mistral's direct API and OpenRouter both list $0.50/1M input and $1.50/1M output. The Coinis retail price includes unified billing, usage dashboards, and no egress overhead across all models in the catalog.

What is the difference between Mistral Large 3 and Mistral Large 2?

Mistral Large 3 (mistral-large-2512) uses a sparse MoE architecture with 675B total parameters and 41B active, a 262K context window, and multimodal image inputs. Mistral Large 2 (mistral-large-2407) is a denser model with a 128K context window and no native image support. On OpenRouter, Mistral Large 2 is priced at $2.00/1M input and $6.00/1M output. approximately 4 times the per-token cost of Mistral Large 3 at the same provider.

Mistral vs. Llama: when should I pick Mistral Large 3 over Llama 3.1 405B?

Choose Mistral Large 3 when EU data governance matters, when you need a 262K context window rather than Llama 3.1 405B's 128K, or when you want Apache 2.0 open weights from a European lab. Llama 3.1 405B is a strong alternative for teams already invested in Meta's ecosystem. Both are competitive on reasoning benchmarks at their respective price points.

Does Mistral Large 3 support tool calling, structured JSON output, and vision inputs?

Yes to all three. Tool/function calling is supported with an observed error rate of approximately 4.09% on the Mistral provider. Structured JSON output via response_format has an observed error rate of approximately 0.45%. Image inputs are accepted via image_url alongside text prompts for multimodal tasks.

Is the Mistral API on Coinis OpenAI-compatible?

Yes. The Coinis endpoint accepts the same request shape as the OpenAI SDK, including messages, tools, response_format, stream, and seed parameters. Point your existing client at POST https://api.app.coinis.com/v1/llm/generate with model set to mistral-large-3 and it works without further changes. See /models/mistral/large-3/api for the full reference.

Should I use mistral-large-latest or pin to mistral-large-2512 in production?

Pin to mistral-large-2512 for production stability. The mistral-large-latest alias always resolves to the newest Large generation, which means a future model release could change response behavior without a code change on your side. Pinning the December 2025 version keeps outputs consistent until you choose to upgrade.

Is my prompt data used to train Mistral models?

Prompt training is explicitly disabled, so your prompts are not used for model training. Per OpenRouter's data policy disclosure, prompt data may be retained by the Mistral provider for up to 30 days for abuse-detection purposes. Check the official Mistral terms at mistral.ai/pricing for the full data handling policy.

Related

Pair it with.

ChatGPT API / GPT-5.4

Model

Claude (Sonnet 4.6 / Opus 4.7 / Haiku 4.5)

Model

DeepSeek V4 Flash

Model

All LLM models

Catalog
Upscale this Edit a frame Add a voiceover

Start free

Your wallet. Every model. One call away.

Start free. 15 tokens a week. No card.

Generate on Coinis

No credit card.

Pricing and capabilities verified 2026-05-26. Read the docs .