Mistral AI: Vibe Chat, Models, Pricing and Benchmark Review

👁 124 reads

Mistral AI is a French artificial intelligence company developing open-weight language models alongside Mistral Vibe, its web and mobile assistant known until May 2026 as Le Chat. On October 6, 2026, the company unveiled its largest architecture yet, Mistral Large 4 (“Le Chonk”, 1.05 trillion parameters), released immediately via API with open weights scheduled for late October. The chat platform is free to use, Pro costs $14.99 per month, and most models can be downloaded and run locally under an Apache 2.0 license. In linguistic benchmark evaluations, Mistral produces grammatically solid copy but trails top commercial peers-Large 4 scored 69.2 points out of 100 versus 80.4 for Claude Sonnet 5.5. Here is how the lineup performs.

What is Mistral AI

Mistral AI was founded in Paris in April 2023 by Arthur Mensch (formerly of Google DeepMind) alongside Guillaume Lample and Timothée Lacroix (both formerly of Meta, where they co-created the original Llama). You can read more about the CEO in my profile on Arthur Mensch. From day one, the team focused on two key principles: open model weights and European data sovereignty, providing on-premise execution or processing within EU-based data centers.

The Mistral ecosystem spans three core products:

• Language models - A portfolio scaling from Ministral 3 3B up to Mistral Large 4 with 1.05 trillion parameters. Most models use the permissive Apache 2.0 license, allowing you to download them from Hugging Face for unrestricted commercial deployment.
• Mistral Vibe - The consumer and workplace conversational assistant accessible in your browser and on mobile, serving as an alternative to ChatGPT. Named Le Chat until May 2026, it features live web search, image generation, voice chat, project workspaces, system memory, and over 100 integrations. In 2026, it expanded to include developer workflows with an agent designed for the terminal and code editor.
• Enterprise platform - Developer API (Mistral Studio), on-premise deployment services, and specialized models covering OCR, speech recognition, content moderation, and document retrieval.

Financial backing: On September 8, 2026, Mistral announced a €3 billion Series D round valuing the company at over €21 billion. Prior strategic rounds, including backing from ASML and discussions with Samsung, are covered in my analyses of Mistral’s valuation and Samsung’s investment. In September, the organization also announced an integration partnership with Mozilla for the Firefox browser assistant alongside opening an engineering hub in Munich.

Mistral Large 4: October 2026 Flagship

Mistral Large 4 is the largest architecture the laboratory has ever deployed and marks the first flagship “Large” upgrade since December 2025. Mistral positions it as the premier open-weight model originating from the US or Europe. Key specifications include:

Mistral tailored the architecture heavily for cybersecurity, financial analysis, legal compliance, and programming: the company reports 93% on Cybench, 82% vulnerability reproduction in the AA Cyber Index, and 61.7% on DeepSWE. When we tested it on multilingual tasks using Polish linguistic checks, it cleared core vocabulary benchmarks (46/46 on grammar quizzes) but showed weaker creative writing fluency than similarly priced alternatives. In thinking mode, it also exhausted its generation budget twice across 13 tasks without returning an answer. Further reading:

• Mistral Large 4: Architecture, Pricing, and Weights - comprehensive launch specifications,
• Mistral Large 4: 13-Prompt Evaluation - test results, reasoning quirks, and API costs,
• Mistral Large 4 vs Claude, GPT, Gemini, and DeepSeek - head-to-head pricing and score matrix.

Running Large 4 locally on standard consumer equipment is unfeasible: 1.05 trillion parameters require roughly 525 GB of memory even under aggressive 4-bit quantization. Local self-hosting requires the smaller releases detailed below.

How to Get Started with Mistral

Mistral Vibe Web and Mobile Chat

1. Navigate to chat.mistral.ai or download the “Vibe by Mistral” application from the Apple App Store or Google Play Store.
2. Register an account. The Free plan does not require credit card details.
3. The interface automatically matches your browser or mobile system language.
4. Start chatting. If you require consistent responses in a specific foreign language, state that upfront in the prompt or in your workspace project instructions.
5. For complex research, open a dedicated project and attach source files-Vibe retains contextual memory across individual sessions.

Running Mistral Models Locally

1. Download and set up Ollama or LM Studio.
2. Select a model suited to your hardware VRAM: Ministral 3 8B (6 GB in Ollama, requires an 8 GB GPU), Ministral 3 14B (9.1 GB, requires a 12 GB GPU), or Mistral Small 3.2 24B (14 GB, requires a 16-24 GB GPU).
3. Run ollama run ministral-3:8b in your terminal. Ollama pulls the weights on first run, after which it starts instantly.
4. Chat directly in your console or hook the local API endpoint (http://localhost:11434) into your client tools.

Higher-parameter architectures (Small 4 with 119B parameters, Medium 3.5, Large 3, and Large 4) will not fit on consumer cards. They require multi-card H100 clusters or external API access.

Mistral Model Overview

Important details regarding the model index:

• Large 4 reclaims the performance crown. From April through October, Mistral Medium 3.5 led the lineup, unifying the experimental Magistral (reasoning) and Devstral (programming) branches. Large 4 outperforms Medium 3.5 according to Artificial Analysis while cutting output generation costs ($4.18 versus $7.50 per million tokens). In our multilingual evaluations, both achieved comparable marks: 69.2 vs 68.0 points.
• Small 4 is an efficient MoE. Despite its 119-billion-parameter total capacity, only a small fraction routes through each token pass, resulting in rapid inference and low serving costs. It supports dynamic reasoning effort parameters (reasoning_effort).
• Third-party models on Mistral Studio. Mistral’s developer documentation lists GLM 5.3 alongside native models. As explored in our piece on European ChatGPT alternatives, commercial platforms operating out of the EU occasionally route queries through foundational models developed abroad.
• Foundational lineage. Numerous community projects are built directly on top of Mistral base models, such as the open-source multilingual model Bielik (7B v0.1 released in April 2024). Explore our guide on Bielik for more on this ecosystem.

Mistral Pricing Breakdown

Pricing reflects standard USD rates from the mistral.ai portal. Full developer billing documentation, usage costs per token for every architecture, and invoice records from our evaluations can be found in our dedicated report on Mistral pricing.

Language Capabilities and Writing Nuance

Mistral understands complex cross-lingual inputs reliably and crafts grammatically correct output, but struggles to achieve the organic phrasing delivered by proprietary leaders. While Mistral models were trained on expansive multilingual datasets spanning major European languages, specific localized quirks emerged during testing:

• Vocative declension and honorifics. When tested in Slavic languages like Polish, Mistral Medium 3.5 failed proper vocative casing in formal correspondence, addressing “Tomasz” without correct inflection.
• Unsolicited emoji usage. Three out of four cloud-hosted Mistral models appended decorative emojis to serious explanatory outputs aimed at young audiences without being asked to do so.
• Grammatical gender consistency. Ministral 3 8B mistakenly swapped first-person grammatical gender when responding to a prompt signed by an unambiguously female sender, while also fabricating details absent from the user prompt.

Across our comprehensive 13-task writing benchmark assessed under blind evaluation, Mistral Large 4 reached 69.2 points with active reasoning and 65.6 without, Mistral Medium 3.5 hit 68.0 points, and the self-hosted Mistral Small 3.2 24B scored 53.9 out of 100-trailing open alternatives like Gemma 4 12B and Bielik 11B. Check the complete data tables in the multilingual AI writing rankings and our study of 7 local models.

Hands-On Evaluation: 5 Complex Prompts

On October 3, 2026, we tested four Mistral API models alongside Ministral 3 8B deployed locally via Ollama against five standardized tasks (we tested it in Polish: strict word-budget summarization, formal administrative correspondence, a 10-error copyediting test, conceptual simplification for children, and idiomatic translation). The pool included reference outputs from models like Gemma 4, Llama, and Muse Glimmer. Output quality was evaluated under blind conditions by Claude Opus 5.5 on a 1-10 scale.

For comparison, the top-performing model across this run was Muse Glimmer 30B (8.67), followed by local Gemma 4 12B (8.33) and Bielik 11B (7.60). Mistral’s premier offerings roughly matched Bielik, while smaller edge options struggled.

Key observations from the test outputs:

• Medium 3.5 proved best at adapting English idioms (e.g., rendering “cut to the chase” and “cost an arm and a leg” into idiomatic equivalents), yet inserted grammatical typos into its proofreading exercise and erred on formal vocative greetings.
• Large 3 and Small 4 inverted meaning when translating the phrase “missed the deadline by a hair”, misinterpreting it as “barely finished on time”.
• Three cloud variants concluded an educational prompt for an 8-year-old with decorative emojis, and Small 4 scientifically claimed atmospheric color shifts were caused by “water droplets and pollen”.
• Ministral 3 8Bgenerated non-existent vocabulary ([corrupted lettering in the original Polish test]) locally, while the cloud API instance hallucinated a fictional discussion with a delivery courier and swapped grammatical gender.

Complete prompt transcripts, citations, and evaluation logs are cataloged in our deep dive on open model benchmark comparisons.

Mistral Compared to Market Alternatives

Assessments of stylistic fluency stem from ongoing direct usage across standard tasks and our benchmark suites. If your primary objective is flawless non-English creative writing without editing passes, Mistral is not the default choice. However, if European legal jurisdiction, GDPR compliance, or self-hosted deployment flexibility are paramount, it stands among the industry’s front-runners alongside Llama and specialized models like Bielik.

Who is Mistral Built For

• Organizations with strict EU data residency requirements - Direct EU corporate accountability, on-premise infrastructure delivery, and fully exportable model weights.
• Software engineers - Highly competitive API pricing (Small 4 at $0.15 input), Apache 2.0 permissive licensing, and integrated terminal coding tools in Pro.
• Enthusiasts with 8-24 GB VRAM GPUs - Ministral 3 8B and Mistral Small 3.2 24B offer broad optimization out of the box in tools like Ollama and LM Studio.
• Not ideal for users demanding hands-off multilingual copywriting that requires zero manual proofreading-Claude and ChatGPT maintain a noticeable edge here.

Frequently Asked Questions

Is Mistral AI free to use?

Mistral Vibe includes a complimentary tier with usage rate limits. Models distributed under Apache 2.0 (such as Ministral 3, Small 4, and Large 3) are completely free to download and use, including in commercial applications. Subscribing to the Pro plan ($14.99 per month) or querying the developer API incur standard usage fees.

What happened to Le Chat?

In May 2026, Le Chat was rebranded to Mistral Vibe. The chat.mistral.ai portal remains active, with user histories, active sessions, and subscriptions carrying over seamlessly. You can read our background overview in Le Chat (now Mistral Vibe).

Does Mistral handle multilingual prompts?

Yes. The platform provides a localized UI across multiple languages, and its underlying weights are trained on broad multilingual corpora. While it comprehends complex prompts reliably, our evaluations revealed occasional minor slips in formal declensions and localized idioms.

Which Mistral model delivers the best performance?

As of October 6, 2026, the flagship model is Mistral Large 4 (scoring 38.4 on the Artificial Analysis Intelligence Index and 69.2 in our laboratory benchmark). It succeeds Mistral Medium 3.5 from April 2026. In our focused 5-prompt evaluation prior to Large 4’s debut, the more affordable Mistral Large 3 marginally outperformed Medium 3.5 (7.67 vs 7.13 out of 10) due to accurate copyediting and letter generation, though Medium 3.5 took first place in translation nuances.

What does Mistral API access cost?

API rates start at $0.10 per million tokens (Ministral 3 3B) up to $1.50 input / $7.50 output per million tokens (Medium 3.5). Mistral Large 4 is priced at $1.36 input / $4.18 output, with a 50% promotional rate applied at launch. Our five-prompt test suite incurred total charges between $0.0003 (Ministral 3 8B) and $0.0061 (Medium 3.5).

What is Mistral Large 4 “Le Chonk”?

The company’s largest MoE architecture, introduced October 6, 2026: 1.05 trillion parameters total, 49 billion active per token pass, multimodal image recognition, and a 1M token context capacity. “Le Chonk” serves as an internal lighthearted reference to Mistral’s feline mascot motifs. API access opened immediately, with public open weights slated before the end of October. Read more in our guide to Mistral Large 4.

Can you run Mistral models locally on your own computer?

Yes. Ministral 3 8B requires approximately 6 GB in Ollama and executes comfortably on an 8 GB graphics card; Mistral Small 3.2 24B requires roughly 14 GB of memory. You can initiate the 8B model via the command: ollama run ministral-3:8b.

Mistral vs ChatGPT: How do they compare?

ChatGPT produces more natural copy across global languages and offers lower-tier subscription entry points. Mistral counters with European governance, guaranteed EU data handling paths, and completely downloadable weights for local execution. For personal creative tasks, ChatGPT often has an edge; for organizations handling sensitive records, Mistral or local open-weight pipelines provide superior data control.

Is Mistral AI secure and privacy-compliant?

Mistral is a European entity subject to GDPR regulations. Its commercial tiers offer contractual data residency within the EU alongside on-premise hardware deployments. As with any public cloud solution, you should avoid entering confidential internal credentials or unencrypted personal records into standard free chat interfaces.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Sources and Verification Dates

• Mistral: Introducing Mistral Large 4 (Oct 6, 2026) and official model documentation
• Mistral AI Press Releases (Series D announcement Sept 8; Firefox collaboration Sept 16; Munich office opening Sept 28, 2026)
• Mistral Vibe Pricing Schedules
• Mistral Model Documentation (versioning, licensing terms)
• Mistral Small 4 Model Card (119B parameters, 128 routing experts)
• Wikipedia: Mistral Vibe (rebranding historical records, May 2026)
• Vibe application release notes on Apple App Store
• Ministral 3 on Ollama Library (quantized footprint benchmarks)
• Hands-on testing data: 5-prompt evaluation (OpenRouter and Ollama 0.35.1 on an RTX 4090, Oct 3) and 13-prompt benchmark for Large 4 (OpenRouter, Oct 7)

Verified on October 7, 2026 (Large 4 specifications), with general platform features confirmed October 3, 2026.

Kinetyra