From research model to heavyweight: the Kimi K2 series

Kimi K2: the technical foundation

The original Kimi K2 version was introduced in July 2025 and described in detail in a technical report (“Kimi K2: Open Agentic Intelligence”). Technically, it is a so-called Mixture-of-Experts (MoE) model: instead of a single, continuously active neural network, the model consists of many specialized sub-networks (“experts”), of which only a small portion is actually activated per request. For Kimi K2, according to Moonshot AI, there are a total of around one trillion parameters, of which around 32 billion are actively used for any single request. The model is spread across 384 experts, of which 8 are specifically selected per text segment, supplemented by an always-active “shared” expert.

This structure explains why Kimi K2, despite its enormous overall size, can be operated comparatively economically: the entire trillion parameters do not need to be computed for every request, only a small, selected part. According to Moonshot AI, the model was trained using a specially developed optimization method (MuonClip), intended to improve training stability for such large MoE models. The original base version offered a context window of 128,000 tokens; the Instruct version (for dialogue and instructions) released at the same time was expanded to 256,000 tokens.

Kimi K2 Thinking: thinking and acting in one model

In autumn 2025, according to consistent reports from November 2025, Moonshot AI introduced “Kimi K2 Thinking,” a variant that additionally incorporates extensive, multi-step thinking before the actual answer and, while still in the thinking process, can already call on tools (such as a web search or a calculator). According to available benchmark figures, the model achieved a score of around 71 percent on the well-known programming test “SWE-Bench Verified” and, in test scenarios, was able to independently carry out several hundred consecutive tool calls to solve complex, lengthy tasks. This deliberately positioned Moonshot AI in the field of so-called agent models, which don’t just answer but independently plan and execute multi-step tasks.

Further development in 2026: K2.5, K2.6 and K3

Over the course of 2026, Moonshot AI released several further tiers. According to consistent, though not always uniform, sources, K2.5 appeared, followed in April 2026 by K2.6 with a context window expanded to 256,000 tokens and improved programming capabilities, followed by a variant specialized in programming tasks called K2.7 Code. Finally, in July 2026, Kimi K3 followed, which, according to the manufacturer, comes with a context window enlarged to around 1 million tokens and native image processing (vision) – according to Moonshot AI, four times the size of K2.6’s context window. Various sources give differing figures for K3’s exact parameter size (some in the range of roughly two to three trillion total parameters); a single, officially definitive figure cannot currently be confirmed with complete certainty, which is why we deliberately speak of an order of magnitude rather than an exact number here.

What makes the Kimi models technically and strategically distinctive?

Open weights with a commercial condition

Moonshot AI releases the model weights for K2 and its successors under a self-formulated, modified MIT license. At its core, this is a very permissive open-source license that also allows commercial use. One special feature: anyone operating a product or service based on the models with more than 100 million monthly active users or more than 20 million US dollars in monthly revenue must visibly display the notice “Kimi K2” in their own interface. For the vast majority of smaller and medium-sized use cases – including in the real estate industry – this clause practically never applies, but it shows that “open” at Kimi is not meant entirely unconditionally.

Agentic capabilities and a large context window

The strategic focus of the Kimi models is clearly on so-called agentic capabilities: the ability to independently handle multi-step tasks with tool use, web access and code execution, rather than just answering individual questions. Combined with a very large context window – around one million tokens for K3 – this makes it possible to process very extensive document collections in a single pass, such as several hundred pages of contract documents or a larger collection of property documents.

Rule of thumb: the large context window makes it possible to read in a great many documents at once – but this does not replace careful review of the output, since even a large context window does not protect against errors or omissions in the result.

Practical examples of use

1. Automated summarizing of large file collections

A case worker at a property management company could have several hundred utility bill statements or minutes from owners’ meetings read in via the API in a single pass, in order to automatically flag anomalies – such as unusually high cost items. The large context window of K2.6 or K3 makes this technically possible without first having to split the documents into many small chunks.

2. Drafting and laying out a listing exposé

In agent mode, a real estate agent can have a first, multi-page draft version of a property exposé created, including a simple web presentation, which is then reviewed editorially by hand and cross-checked against the actual, legally binding property data.

3. Programming support for internal tools

Since the Kimi models perform particularly strongly on programming benchmarks, some IT departments at real estate companies use Kimi to help build small internal automation scripts, for example to reconcile rent lists between two software systems.

4. Research and market monitoring

Via the web-browsing feature of agent mode, a model can be used to continuously monitor publicly accessible market data, for example to summarize interest rate trends or publicly accessible real estate market reports – it’s important here to work exclusively with publicly available, non-personal sources.

5. Multilingual communication

For international real estate portfolios, a financial case worker can use Kimi to translate standard letters or summaries between several languages – though the restriction on personal content described above still applies here as well.

Rule of thumb: the larger the volume of data a model processes in one pass, the more important the question becomes of exactly where that data resides during processing – performance and data protection are two separate questions, and both need to be answered.

FAQ about the Kimi models

What is the difference between Kimi K2 and Kimi K2 Thinking?

Kimi K2 is the original base model; Kimi K2 Thinking is a variant introduced in November 2025 that incorporates more extensive, multi-step thinking with tool use before answering, and thereby performs better on complex tasks.

How large is the context window of the current models?

The models in the K2 series generally offer 256,000 tokens of context; Kimi K3, introduced in July 2026, is stated to have around one million tokens.

Are the Kimi models really open source?

The model weights are openly accessible and released under a modified, very permissive MIT license. One special feature is the attribution requirement for very large commercial applications (over 100 million users or 20 million US dollars in monthly revenue).

Are the models suited to programming tasks?

Yes, the Kimi K2 series, and especially the specialized K2.7 Code variant, perform comparatively strongly on public programming benchmarks and are frequently used for coding agents.

Can I run the models myself instead of using the Kimi website?

Yes, since the weights are released openly, they can be run via various cloud providers or on your own hardware. However, this requires technical expertise and sufficient computing capacity, since these are very large models.

Basic information about the company Moonshot AI, pricing and data protection can be found on our provider page. A comparison of all ten AI providers in the cluster is available on the overview page.

The Grok model family at a glance

Grok 3 – the previous generation

Grok 3 was xAI’s main model for a long time and now forms the basis for the free tier. It handles the basic tasks of a chatbot – answering questions, drafting texts, simple research – but offers a considerably smaller context window and less accuracy for complex, multi-step tasks than the newer models.

Grok 4 and Grok 4 Heavy

Grok 4 was the next big step and brought considerably better results in logical reasoning and more complex questions. The Grok 4 Heavy variant goes a step further: for difficult tasks, it uses several parallel thinking processes and compares the results before giving an answer. This makes Grok 4 Heavy slower and considerably more expensive, but also more reliable for tricky tasks – such as mathematical calculations or multi-step analyses.

Grok 4.1 Fast – fast and inexpensive

For tasks where speed and low cost matter most – such as automatically answering many short inquiries in an app – xAI offers Grok 4.1 Fast, a stripped-down but very fast and inexpensive variant. Its depth of content is lower, but that is usually sufficient for simple, repetitive tasks.

Grok 4.20 – a huge context window and multi-agent technology

With Grok 4.20, xAI introduced a model in early 2026 that internally works with several specialized “roles” that cross-check each other before an answer is given – put simply: one role coordinates, one checks facts, one assesses technical details, one provides creative phrasing. For users, this mainly shows up as more reliable answers for longer, more complicated queries. In addition, there is a very large context window of around 2 million text units (tokens) – equivalent to several thousand pages of text that the model can keep “in mind” at once, for example for analyzing longer contracts or extensive data collections.

Grok 4.5 – the current focus on complex and technical tasks

Grok 4.5, released in July 2026, is currently xAI’s most demanding available model. It was specifically trained for longer, demanding work sessions – originally mainly for software development, but now also for other complex, multi-step tasks such as comprehensive research or the analysis of larger data sets. Its context window is around 500,000 tokens, so smaller than Grok 4.20’s, but the model works more precisely on difficult reasoning tasks. Via the programming interface, Grok 4.5 costs around 2 euros per million input text units and around 6 euros per million output text units.

Grok 5 – what is known so far

A successor model with the working name Grok 5 is, according to xAI, in the training phase on an expanded Colossus data center. There is no official release date as of August 2026; according to rumors, the model is expected to be considerably larger than Grok 4 and to be introduced later in 2026. Until the official announcement, Grok 4.5 remains the most powerful generally available model.

Rule of thumb: the higher the version number, the more powerful and usually also more expensive the model – but not every task needs the strongest model. For simple everyday questions, the fast, inexpensive variant is often enough.

Comparison table of the Grok versions

Real-time data access via X

One feature shared by all current Grok models is direct access to public posts on the platform X. This allows Grok to react to events that are only a few minutes old – a clear difference from models whose knowledge is limited to a certain training cutoff date. The downside: posts on social media are not automatically verified or reliable, which is why answers based on current X posts should always be read with a healthy dose of caution, especially for controversial or unclear topics.

Image and video generation with Aurora and Grok Imagine

For generating images, xAI uses its own system called Aurora, now marketed under the product name Grok Imagine. It generates images from text descriptions and, by now, also short videos a few seconds long in various formats. In the paid SuperGrok tier, the full feature set of Grok Imagine is included; the free tier offers only a limited basic version. Via the programming interface, video generation is billed by the number of seconds generated, which makes it comparatively inexpensive for short clips.

Examples of use from everyday life and business

Example 1 – everyday life: A user is planning a trip and has Grok summarize, based on current posts on X, what the weather and traffic conditions currently look like at the destination – something a model without real-time access cannot deliver.

Example 2 – small business: A case worker at a small trades business uses Grok 4.1 Fast to pre-sort incoming customer inquiries by email and draft reply suggestions, which are then reviewed by a human.

Example 3 – real estate and finance: A real estate office wants to get a first impression of the current sentiment around a particular location – for example, whether there has recently been more discussion about rising rents or a new construction project in a city. Thanks to its access to current X posts, Grok can summarize such discussions and provide initial clues. Importantly: such a sentiment analysis does not replace a robust market analysis based on official data, but merely provides a quick, supplementary impression.

Example 4 – reviewing a long document: A case worker uploads a longer lease or an extensive tender document into Grok 4.20 and has the key clauses, deadlines and figures summarized – thanks to the large context window, the model can take the entire text into account at once, instead of having to break it into sections.

Example 5 – a technical task: A developer uses Grok 4.5 to analyze an existing software application across several files and get suggestions for improving the program code – a task the model was specifically trained for.

Rule of thumb: the longer and more branched the task, the more important the context window becomes – for short everyday questions, on the other hand, it hardly matters.

Practical example: estimating the cost of the programming interface

A real estate office wants to have around 50 property descriptions automatically drafted per month with Grok 4.5. Each description requires an estimated 1,000 tokens of input (property data, conditions) and 1,500 tokens of output (finished text).

• Input: 50 x 1,000 = 50,000 tokens, at 2 euros per million tokens that comes to around 0.10 euros
• Output: 50 x 1,500 = 75,000 tokens, at 6 euros per million tokens that comes to around 0.45 euros
• Total cost per month: around 0.55 euros

This calculation example shows: the pure text costs via the programming interface are very low for manageable text volumes – the actual costs usually come from developing and maintaining the integration, not from the individual requests themselves.

Frequently asked questions (FAQ)

Which Grok model should I use as a beginner?

For getting started, the standard model in the free or the cheapest paid tier is generally sufficient. Switching to more powerful models is only worthwhile once you hit limits with complex or long tasks.

What does “context window” mean in simple terms?

The context window describes how much text a model can keep “in mind” at once. A large context window is important if you want to have long documents analyzed at once, but it hardly matters for short questions.

Is Grok 4.20 the same as Grok 4?

No. Grok 4.20 is its own, later model generation with a considerably larger context window and an internal multi-role check that was not yet present in Grok 4.

When is Grok 5 coming?

There is no official date as of August 2026. xAI has confirmed that a successor model is in the works, without naming an exact date.

Can I get reliable market data for real estate decisions with Grok?

Grok can summarize current discussions and sentiment, but it does not replace verified official data or a professional market analysis. It is best suited as a quick supplement, not as the sole basis for a decision.

An overview of the company xAI itself, its pricing tiers and the data protection assessment is available on our xAI provider page. Anyone wanting to see the other major AI providers compared instead will find the overview on the AI models comparison page.

Why “Copilot” is not just “Copilot”

Microsoft has deliberately turned the name Copilot into an umbrella brand, similar to how other companies do with brands for entire product families. Behind the surface, however, the variants differ considerably: in the data they have access to, in the language model used, and in the security and privacy commitments that apply to each.

Copilot in Windows

The variant integrated into the taskbar is primarily a general-purpose chat for everyday questions, research and simple writing tasks. It has no direct access to internal company documents or emails and is therefore functionally closer to a classic chatbot than a genuine office assistant.

Microsoft 365 Copilot (Business and Enterprise)

This variant is the actual business version. It is connected, via the so-called Microsoft Graph interface, to the documents, emails, calendar entries and chats that the respective user already has access to. This allows it, for example, to create a summary that is actually based on a team’s real project files, rather than merely reflecting general world knowledge. This connection is the decisive technical difference from all other Copilot variants.

Microsoft 365 Premium (successor to Copilot Pro)

The tier intended for private users and self-employed people brings full Copilot features into personal Office applications, but without the deep connection to a company network with many users that the Business variant offers. It is aimed at individuals, not teams.

GitHub Copilot: its own product line

For completeness, it should be mentioned that GitHub Copilot is a standalone product for software developers that assists with writing program code. It shares only the brand name and part of the underlying technology with the other Copilot variants, but is aimed at a completely different target group and is mentioned here only for context.

Which AI models are really behind it

For a long time, practically the entire Copilot family ran exclusively on models from OpenAI. That changed in 2026. In July 2026, Microsoft and OpenAI announced that the new model GPT-5.6 would become the preferred model for Microsoft 365 Copilot in Word, Excel, PowerPoint and the new Copilot workspaces. The model is specifically optimized for knowledge work and is meant to work more efficiently than its predecessors on more complex, multi-step tasks.

At the same time, Microsoft has broadened its strategy and, since 2025, has also offered models from Anthropic in Microsoft 365 Copilot. In practice, administrators in many enterprise environments can now select or restrict which model is used for which application. This is an important difference from pure single-model providers: technically, Copilot works more like a switchboard that mediates between several models in the background, rather than like a single, unchanging language model.

Anyone who asks “which model does Copilot use” no longer gets a simple answer in 2026. The more accurate question is: which model is Copilot currently using for this particular task, and can that be changed in the settings.

Comparison table of the Copilot variants

Practical examples of use from everyday office life

Example 1: Handling an email flood in Outlook

A case worker returns from a two-week absence and finds over 200 unread emails. Copilot summarizes long threads, flags emails requiring direct action, and suggests ready-made reply drafts for routine inquiries. Instead of opening every message individually, the employee gets an overview within a few minutes.

Example 2: Meeting summary in Teams

After a one-hour project meeting, Copilot automatically creates a summary with the most important decisions, open questions, and a list of concrete tasks including the person responsible for each. Anyone who couldn’t attend doesn’t have to watch a recording, but can read the summary in a few minutes.

Example 3: Building a yield spreadsheet in Excel

An example with a real estate angle: a property manager wants to calculate the gross rental yield for several possible purchase prices for an apartment building. He describes to Copilot in Excel, in plain language, which columns he needs, such as purchase price, annual rent, operating costs and the resulting yield in percent. Copilot builds a basic structure with the appropriate formulas from this, which the manager then adapts to the actual figures for his property. Importantly: Copilot provides a usable calculation template, but the substantive review of the formulas and the values used remains the manager’s own responsibility.

Example 4: Revising a listing exposé text in Word

A real estate agency has written a factual but linguistically very dry property description for a listing exposé. Copilot in Word can, on request, give the text a friendlier, more readable tone, suggest shorter sentences, or remove repetition, without changing the actual facts about the property. The human remains responsible for the factual accuracy of the information; Copilot merely handles the linguistic fine-tuning.

Example 5: A first slide outline in PowerPoint

For an internal presentation on quarterly results, Copilot delivers a first slide outline, complete with suggestions for suitable chart types, based on a short bullet-point text. This mainly saves the tedious first half hour of structuring; the actual substantive elaboration remains manual work.

What the model switch concretely means for users

For daily work in Word, Excel or Outlook, most users hardly notice the switch from one model to the next, since Microsoft largely controls model selection automatically in the background. The topic becomes relevant mainly for IT managers at companies, who can determine in the admin settings which models are even permitted for their organization at all, for example for data protection reasons. For private users, generally only the choice of subscription tier remains, not the individual model selection.

Frequently asked questions about the Copilot models

Is GPT-5.6 available in every Copilot variant?

No. GPT-5.6 is used mainly in Microsoft 365 Copilot and in Microsoft 365 Premium. The free Windows variant sometimes uses simplified or older model versions.

Can I, as a user, choose which model Copilot uses myself?

In enterprise environments, administrators can restrict or enable model selection. Private users generally have no direct model selection, only the choice of subscription tier.

Why does Microsoft use Anthropic models alongside OpenAI?

Since 2025, Microsoft has deliberately pursued a multi-model-provider strategy, so it can deploy suitable models for different tasks and not be dependent on a single provider.

Is GitHub Copilot the same as Microsoft 365 Copilot?

No, GitHub Copilot is a standalone product for software development with its own models and its own pricing structure, sharing only the brand name with the office version.

Are Copilot’s calculation results in Excel reliable?

Copilot provides a good basic structure and suitable formula suggestions, but the substantive check of the values and formulas must still be carried out by a human, especially for business-relevant calculations.

An overview of the individual pricing tiers, data protection questions, and the general assessment of Microsoft Copilot is available on the provider page Microsoft Copilot overview. Anyone wanting to see the other eight providers compared instead will find the overview on the main page comparison of all AI models.

The basic principle: web search plus language model instead of an answer from memory

Why a pure language model alone is not enough

A large language model, such as the kind behind classic chatbots, was trained with huge amounts of text at a certain point in time. After that, it has a great deal of general knowledge, but no automatic access to new events, current figures, or freshly published studies. If you ask such a model about the current base interest rate or a change in the law passed yesterday, it can either honestly say it doesn’t know, or – in the worse case – invent something false that sounds plausible. This is exactly the problem Perplexity solves with a two-stage process.

The interplay: retrieval meets text generation

In the first stage, so-called retrieval, Perplexity searches current web pages, news articles, trade portals and, in some cases, scientific databases for content matching the question asked. These results are filtered, sorted by relevance and trustworthiness, and passed on as a basis to a language model. Only in the second stage does the language model condense these results into an understandable, coherent answer and attach a footnote to every statement pointing to the respective source. You can picture this like a very fast research assistant who first reads ten trade articles and then dictates a summary with source citations – except this process takes only a few seconds.

It’s not that the language model “knows” the answer to a current question – it reads it out of freshly found sources in real time and translates it into understandable language.

What models are specifically behind this?

Proprietary Sonar models for standard queries

For most everyday queries, Perplexity uses its own, internally developed models from the so-called Sonar series. These are specially trained to efficiently evaluate web search results and quickly turn them into a clean, source-backed answer. The advantage of proprietary models lies mainly in speed and cost efficiency – for simple factual questions, the most elaborate available model doesn’t need to be called on.

Access to external top-tier models depending on subscription

For more complex questions, longer analyses, or on explicit request, Perplexity additionally draws on external language models from major providers – depending on availability and subscription tier, for example current models from OpenAI’s (GPT), Anthropic’s (Claude) and Google’s (Gemini) lineups. Perplexity itself does not develop these models but integrates them via programming interfaces and combines them with its own search technology. Pro users can often specifically choose in the advanced settings which model should formulate an answer, while the free version makes this choice automatically in the background.

Model Council: comparing several top-tier models at once

For Max subscribers, there is a special feature, often called “Model Council”: Perplexity processes the same question in parallel with several leading language models at once, compares the results, and summarizes where the models agree, where they diverge, and which aspects each model worked out particularly well. For the average user, this is usually unnecessary effort, but for people who want to thoroughly back up a truly important and complex decision, the model comparison can provide valuable additional perspectives.

Deep Research: multi-step, self-directed research

In addition, there is a mode called Deep Research, in which Perplexity does not just perform a single search but independently plans several search steps one after another: first an overview is established, then specific sub-questions are researched further, and in the end a longer, structured report with many source references is produced. This mode takes considerably longer than a normal query, but delivers more thoroughly grounded results for more complex topics such as market analyses.

Modes and tiers compared

Examples of use from everyday life and work

Example 1: A quick factual question in everyday life

A user wants to know the current inflation rate in Germany. The standard search delivers a short answer within a few seconds, including the current figure along with a reference to the corresponding release from the Federal Statistical Office – considerably faster than clicking through several web pages yourself.

Example 2: Researching current mortgage rates with source citations

A case worker in real estate financing wants to create a rough overview of the current trend in mortgage rates for an internal note. He asks about the approximate interest rate level for ten-year mortgage loans and has the underlying sources displayed, such as current reports from financial portals or interest rate statistics. It’s important here: Perplexity provides good initial orientation with traceable evidence, but does not replace binding financing advice or a same-day bank rate inquiry. Still, the benefit is real for a quick, source-based assessment before a consultation.

Example 3: Summarizing a longer market report or lease

A property manager who uploads a long market report or a multi-page lease as a PDF into a Space can then ask targeted questions about it, such as the key takeaways, notice periods, or specific figures. Perplexity draws on both the uploaded document and, additionally, the current web search to put statements into context – but this does not replace a legal review by a professional.

Example 4: Preparing for a professional discussion with several perspectives

Before an important decision, such as a strategic direction for your own business, the Model Council mode can be helpful: three different top-tier models process the same question, and the summary shows where the assessments agree and where they diverge. This does not replace professional advice, but it does provide a broader viewpoint than a single AI answer.

Example 5: Quick translation and contextualization of foreign-language sources

Since many trade articles and studies are only available in English, Perplexity can incorporate foreign-language sources directly into a German-language summary, including a reference to the original. This often saves the detour of using a separate translation tool.

What does this technology mean for the reliability of the answers?

The combination of web search and language model considerably reduces the risk of freely invented statements, since ideally every claim can be traced back to an actual source. However, an error is not entirely ruled out: the underlying web sources themselves can also be outdated, incomplete, or simply wrong, and a language model can lose nuances when summarizing. The golden rule therefore remains: source citations are an enormous help when verifying, but they are no substitute for your own critical eye, especially for decisions with business or financial relevance.

The more important the decision, the more worthwhile it is to look at the original source behind the footnote – not just at the AI summary.

Frequently asked questions (FAQ)

Which AI model does Perplexity actually use?

There is no single fixed model. Perplexity combines its own Sonar models with external top-tier models from various providers and automatically or, on request, selects the appropriate combination depending on the query, mode and subscription tier.

What is the difference between Pro Search and Deep Research?

Pro Search answers a question in a single pass, while Deep Research independently plans several search steps one after another and, in the end, delivers a more detailed, structured report.

Is the Model Council mode worthwhile for private individuals?

Generally not for simple everyday questions, since the extra effort brings hardly any added value. For truly important, consequential questions, however, viewing the topic from several model perspectives can be worthwhile.

Can I blindly rely on the source citations?

No. The footnotes make verification considerably easier, but they do not replace your own assessment of the source, especially when a decision has financial or legal consequences.

Why does Perplexity sometimes give different answers to the same question?

Since different models and search results can be used in the background depending on load, subscription tier, and sources found, slightly different phrasing or emphasis is normal.

An overview of Perplexity AI as a whole product, its pricing, and its data protection assessment is available on the provider overview page: Perplexity AI: the AI search engine with source citations, overview. A comparison with the other major AI providers in the cluster can be found on the overview page: AI models compared.

From Parisian startup to European AI champion

The focus here is on Mistral AI as a company, its product Le Chat (renamed Mistral Vibe in 2026), the model family from Large to Small, the current pricing models, and the question of who benefits most from this provider – placed in context within our overview of AI models compared.

• Mistral AI website
• Open Mistral Vibe directly

Founded by former Meta and Google researchers

Mistral AI was founded in Paris in April 2023. The three founders – Arthur Mensch, Guillaume Lample and Timothée Lacroix – did not come out of nowhere: they had previously worked on large language models at Meta AI (Facebook’s AI research lab) and at DeepMind (Google’s AI subsidiary). This background is no coincidence; it explains why Mistral was technically able to compete on equal footing with the US providers from the very beginning, even though the company is many times smaller than OpenAI, Google or Meta.

The mission: Europe’s own answer to US AI dominance

Mistral explicitly sees itself as a European counterweight to technological dependence on American cloud and AI corporations. The company is often described as “European digital sovereignty in action”: France and the EU also support Mistral politically, because a capable AI company headquartered and legally based in the EU is, for many public bodies, ministries and regulated industries, simply the only option that satisfies European data protection law without detours.

Rule of thumb: Mistral AI is no small side player, but currently the most important AI company headquartered and running data centers within the European Union.

The product: from Le Chat to Mistral Vibe

For years, Mistral’s chat application was simply called Le Chat (“the chat”) and could be used as an app and in the browser – as a direct competitor to ChatGPT. In May 2026, Mistral renamed its assistant and developed it further into Mistral Vibe. Existing customers automatically kept all their previous conversations, settings and subscriptions – only the name and the feature set changed.

What Mistral Vibe (formerly Le Chat) can do today

• Chat function: classic conversation with the AI model, as known from other providers.
• Web search: current information from the internet is built directly into the answer.
• Image generation: creating graphics and illustrations from text descriptions.
• Document analysis: PDFs, contracts or spreadsheets can be uploaded and summarized.
• Work Mode: a connection to Google Workspace, Outlook and Slack, through which the AI manages appointments, emails and tasks directly.
• Code Mode: a connection to GitHub, which automates simple programming tasks up to finished “pull requests” (prepared code changes submitted for review).

For most private users and smaller offices, the first four points are the most relevant. Work Mode and Code Mode are aimed more at companies with their own IT department.

The Mistral model family: a model for every task

Anyone who looks more closely at Mistral AI quickly comes across a whole series of different model names: Large, Medium, Small, Codestral, Devstral, Magistral, Ministral, Pixtral. At first this seems confusing to laypeople. In fact, however, there is a clear system behind this variety: Mistral offers models of different sizes and different specializations for different use cases – and even makes some of them available for free self-hosting.

Mistral broadly distinguishes between general-purpose language models (for text understanding, summaries, research, everyday questions) and specialized models for specific tasks such as programming, logical reasoning, or understanding images. Anyone using Mistral Vibe in everyday life usually has the selection handled automatically in the background – but anyone deliberately using the models via the programming interface (API) for their own applications should know the differences.

General-purpose language models

• Mistral Large 3: Mistral’s flagship top-tier model, released at the end of 2025. It offers the highest response quality, a very large context window of around 256,000 text units (tokens), and is suited to complex tasks such as long contract analyses or demanding research.
• Mistral Medium 3.5: a mid-range model updated in spring 2026. It delivers a good compromise between speed, cost and quality and is used by many small and medium-sized businesses as an everyday model.
• Mistral Small 4: a compact, very fast and inexpensive model for simple tasks with a high message volume, such as automated short replies in customer service.

Specialized models

• Codestral: specialized in program code, used in many development environments as autocomplete.
• Devstral 2: a large model specialized in complete programming tasks, capable of working independently from the task description to a finished code suggestion.
• Magistral: Mistral’s reasoning model family, specialized in multi-step logical reasoning, such as mathematical or legal questions with several intermediate steps.
• Pixtral: originally the name of Mistral’s image-understanding model, for example for reading scanned documents or photos. The standalone Pixtral brand was discontinued in 2026; the image capabilities are now built directly into Large 3, Medium 3.5 and Small 4.
• Ministral (3B / 8B / 14B): very small, resource-efficient models for use on mobile devices or in applications without a permanent internet connection.

Rule of thumb: bigger is not automatically better suited – for simple, frequently recurring tasks, a small model often delivers a sufficiently good result faster and cheaper than the large flagship model.

Open versus closed models: a unique feature of Mistral

A key difference from many US providers: Mistral releases several of its models as open models (open weight) under the Apache 2.0 license. This means anyone can download these models for free, run them on their own hardware, and even adapt them for their own purposes. The openly available models include, among others, Mistral Small, the compact model “Nemo,” the Ministral 8B variant, the older Mixtral model, and Codestral Mamba.

The large flagship model Mistral Large, as well as the latest Medium versions, on the other hand, remain proprietary, i.e. usable only via the paid API or via Mistral Vibe. For companies with particularly high security requirements – such as public authorities that are not allowed to let any data leave their own data center at all – exactly this combination is interesting: hosting an open model yourself and thereby retaining full control, without having to forgo AI use altogether.

Comparison table of the most important model versions

Pricing models at a glance

Mistral deliberately relies on a simple, transparent pricing model with few tiers. This sets the provider apart from some competitors who work with a confusing variety of subscription options.

Important for context: Mistral’s Pro subscription is considered by trade press to be one of the cheapest subscriptions among the major AI providers that offers full access to a genuinely powerful flagship model. Anyone working with AI daily saves noticeably here compared to some US competitors.

Data protection and GDPR: Mistral’s biggest advantage

The most important difference between Mistral and the well-known US or Chinese AI providers lies not in the raw performance of the models, but in the legal framework. Mistral AI is a French company headquartered in Paris and is therefore fully subject to European law – including the General Data Protection Regulation (GDPR) and the new EU AI Act.

Why this matters concretely for German users

With many US providers, companies and public authorities must check whether a data transfer to the USA is even permissible at all, which adequacy agreement applies, and whether additional contractual clauses are necessary. This review costs time, often requires external data protection consulting, and remains a certain residual legal risk, because the political situation surrounding such agreements has repeatedly changed.

With Mistral, this problem largely disappears: the servers can optionally be located in France or Germany, there is no data transfer to third countries, and no reliance on shaky international agreements. For case workers, municipalities, law firms, medical practices or even property management companies working with particularly sensitive data, this is a tangible practical advantage, not merely a marketing argument.

Rule of thumb: model size alone does not determine whether an AI is suitable for use with sensitive data – what matters is which legal jurisdiction the company and its servers are actually located in.

Strengths and weaknesses of Mistral AI

Who is Mistral AI suited for?

Mistral AI is particularly well suited to anyone for whom data protection in artificial intelligence matters.

• Public authorities and government bodies
• Law firms and medical practices
• HR departments
• Small businesses with customer data
• Self-employed people with tenancy relationships

Anyone looking purely for the most powerful model, for whom data protection is secondary, may find more features with US providers in some cases.

Practical example: cost comparison for a small office

A real estate office with five employees uses Mistral for tenant correspondence and contract summaries.

• Pro subscription: 15 euros per person/month
• Team cost: around 75 euros/month
• Full model for all employees
• Document analysis for leases included
• Data provably stays within the EU

With a comparable US provider, a similar price often applies. In addition, however, the office would have to check on what legal basis tenant data may be transferred to the USA – with Mistral, this step is unnecessary.

Anyone interested in AI for the real estate sector will find a German complement for real-estate-specific questions with Lukinski AI.

Practical examples: which model for which task

Example 1: Summarizing a lease

A case worker at a property management company uploads a twelve-page lease into Mistral Vibe and has the most important clauses on notice periods and utility costs summarized in simple sentences. Because the data is processed in a Mistral data center in Europe, the otherwise necessary review of a third-country transfer is unnecessary – a clear time saving compared to using a US service for the same purpose. Still important: the AI provides a summary for orientation purposes, not legally binding information.

Example 2: Calculation example for customer service automation

An online shop receives around 500 short customer inquiries a day about shipping status and returns. With the compact model Mistral Small 4, processing per 1 million text units costs only a fraction of the price of the large Large 3 model. At around 500 inquiries per day with an average of 300 text units per inquiry and answer, that works out to about 150,000 text units per day, or roughly 4.5 million per month. With the small model, the monthly computing costs for this stay in the low single-digit euro range – with the large flagship model, the same process would be many times more expensive, without the response quality for this simple task being noticeably better.

Example 3: Programming support with Codestral

A small software team uses Codestral directly in its development environment to have program code autocompleted. This saves, on average, a double-digit percentage of writing time for recurring code blocks.

Example 4: Image recognition with Mistral Small 4

An insurance company has damage photos pre-checked automatically using the image capabilities built into Mistral Small 4 (previously marketed as the standalone Pixtral model), in order to spot obvious false claims before a person handles the case’s final review.

Example 5: Multi-step calculation with Magistral

A tax office has Magistral prepare a multi-step example calculation for depreciating a property, in order to write out the individual calculation steps in a way clients can follow – the final review, of course, remains with the tax advisor.

Rule of thumb: the more sensitive or the higher the volume of a task, the more important it becomes to deliberately choose the right model – not every task needs the most expensive flagship model.

How do you choose the right model?

Frequently asked questions about Mistral AI

Is Le Chat now the same as Mistral Vibe?

Yes. Mistral renamed its chat application from Le Chat to Mistral Vibe in May 2026 and expanded it with additional features such as Work Mode and Code Mode. Existing accounts, chat histories and subscriptions were carried over automatically.

Is Mistral AI really GDPR-compliant?

As a French company, Mistral AI is directly subject to the GDPR and the EU AI Act. Servers can be operated in France or Germany, and no data transfer to the USA is required. For legally sound use within your own company, the current Mistral data processing agreement should nevertheless be reviewed.

What does Mistral AI cost compared to other providers?

The free model is enough for occasional use; the Pro subscription costs around 15 euros a month and is considered one of the cheapest full-access options to a top-tier model on the market.

What is the difference between Mistral Large and Mistral Small?

Large is the large, most powerful and most expensive model for complex tasks. Small is considerably more compact, faster, cheaper and intended for simple, high-volume tasks. Small or Medium is entirely sufficient for most everyday questions.

Can I run Mistral models on my own servers for free?

Yes, several models such as Mistral Small, Ministral 8B or Codestral Mamba are openly licensed (Apache 2.0) and can be downloaded and self-hosted for free, provided the appropriate hardware is available – an advantage over almost all closed US providers.

Which Mistral model is suited for summarizing contracts?

For longer contracts, Mistral Large 3 is best suited because of its large context window, since it can fully process even very long documents.

Is Magistral the same as Mistral Large?

No. Magistral is its own model family specialized in multi-step logical reasoning, while Large is a general-purpose language model for broad use cases.

For which professions is Mistral especially worthwhile?

Professions with high data protection needs benefit especially: administration, law firms, medical practices, HR departments and property managers who regularly work with personal data.

A comparison of all ten AI providers in our cluster is available on the overview page AI models compared.