Mistral Large 4 card readers on AIxploria are invited to “discover” Mistral’s newest model in a single paragraph. The directory published the entry at 05:28 UTC on Wednesday 7 October 2026, a day after Mistral released the model as a public preview. It calls the model “a powerful open-source model (1,000 billion parameters)” and tags it as free.
Most of the paragraph checks out. Two of its most visible labels do not, and its rank badge points at another model’s slot. We covered the model itself in our Mistral Large 4 launch report and its place on the leaderboards in our Mistral Large 4 ranking explainer. This article is about the directory listing.
Below we read the Mistral Large 4 card line by line against Mistral’s own announcement and Artificial Analysis data, test the rank badge, look at the alternatives it suggests, and set out what a business should check before acting on any card like it.
Table of contents
- What Mistral Actually Launched on 6 October
- The Mistral Large 4 Card at a Glance
- Mistral Large 4 Card Claims Checked, One by One
- The “Open-Source” Label on the Mistral Large 4 Card
- The Free Tag on the Mistral Large 4 Card
- The Rank Badge on the Mistral Large 4 Card
- The Alternatives Strip on the Mistral Large 4 Card
- Where the Mistral Large 4 Card Has Spread
- What the Mistral Large 4 Card Leaves Out for Businesses
- How to Use a Directory Card Like This One
- Mistral Large 4 Card: Frequently Asked Questions
- References
What Mistral Actually Launched on 6 October
Mistral Large 4 is Mistral AI’s largest model: a mixture-of-experts model with 1 trillion parameters, of which 49 billion are active for each token. Mistral’s own nickname for it is “le Chonk”. The preview runs on Mistral’s API, and Mistral says the weights will be released “by the end of the month”.
The short version of the launch
Mistral’s post calls it “our largest and most capable model to date”. It takes text and images as input and answers in text. Mistral trained it on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centres and serves the preview on the same infrastructure. The list price is $1.36 per million input tokens and $4.18 per million output tokens.
Why a directory card matters
Many people meet a new model through a directory rather than through the vendor’s post. A Mistral Large 4 card is often the first summary a buyer reads, and its labels, such as “open-source” and “Free”, travel into shortlists and slide decks. That makes small errors on a card worth checking.
The Mistral Large 4 Card at a Glance
The Mistral Large 4 card is a short entry with no long review. Its page metadata records a word count of two for the main body, and the visible text is a single blurb plus tags, a badge and buttons. Everything below comes from the page as we downloaded it at 10:33 UTC.
| Field | What the card shows |
|---|---|
| Published | 05:28 UTC, 7 Oct 2026 (edited 05:33) |
| Page slug | mistral-large-4-le-chonk |
| Rank badge | #27 in LLM models |
| Status label | Verified Tool |
| Pricing tag | Free, price 0 |
| Rating | 5 out of 5 from 1 vote |
| Upvotes | 69 |
| Tags | Latest AI, LLM models |
| Similar tools shown | Claude Opus 5.5, GPT-6 Astra, Claude Fable 5.1 |
One vote, 69 upvotes
The five-star score comes from a single vote, which is in the page’s structured data rather than on screen. The 69 upvotes are a separate counter. Neither number says anything about quality five hours after publication. They measure attention in the directory, nothing more.
No review block
Larger AIxploria cards carry a written review with pros, cons and a verdict. This one does not. So the whole Mistral Large 4 card rests on a 38-word blurb and its labels, which is why each label deserves a check.
A one-day lag from launch
Mistral launched on Tuesday 6 October and the card appeared early on Wednesday, a lag of about one day. That is quick by this directory’s standards. We measured one day for Gemini 3.8 Flash, two for Claude Fable 5.1 and three for GPT-6 Astra in earlier checks.
Mistral Large 4 Card Claims Checked, One by One
We split the blurb, the tags and the page metadata into twelve separate claims and checked each against Mistral’s announcement, Artificial Analysis and our two earlier articles. The verdicts use four labels: correct, incomplete, premature and wrong.
| Claim on the card | What the sources say | Verdict |
|---|---|---|
| “open-source model” | Weights due end of October; licence unpublished | Premature |
| 1,000 billion parameters | 1 trillion total, 49 billion active | Correct |
| Understands text and images | Text and images in, text out | Correct |
| Technical drawings and satellite imagery | Both shown in Mistral’s demos | Correct |
| Fixes vulnerable code | 82% on a reproduce-and-patch test | Correct |
| More than 160 languages | “Natively fluent in 160+ languages” | Correct |
| Excellent agent capabilities | 59.9% on AutomationBench, vendor-reported | Incomplete |
| Coding, agents, cyber, finance, law, vision | The verticals Mistral names | Correct |
| Free, price 0 | Preview API is priced per token | Wrong |
| Web-based | API now, self-hosting after weights ship | Incomplete |
| #27 in LLM models | Slot 27 belongs to Claude Sonnet 5 | Wrong |
| Similar tools: three closed models | No open-weight rival in the visible list | Incomplete |
Bars are scaled to the largest group, six correct claims, so each width is the count divided by 6: 3 divided by 6 is 50%, 2 divided by 6 is 33.3% and 1 divided by 6 is 16.7%. The four groups add up to the twelve claims in the table.
Half right is still a fair summary
Six of twelve claims are fully correct, and the technical ones hold up best. Parameter count, input types, vision use cases, the cyber result and the language count all match Mistral’s post. The problems sit in the labels a buyer is most likely to copy: licence, price and rank.
Vendor claims, repeated faithfully
Several “correct” verdicts only mean the card repeats Mistral accurately. The satellite and engineering-drawing skills, the 160 languages and the cyber score are Mistral’s own claims. Independent tests of most of them do not exist yet, a caveat the Mistral Large 4 card does not mention.
The "Open-Source" Label on the Mistral Large 4 Card
The first adjective on the Mistral Large 4 card is the most consequential. “Open-source” tells a reader they can download, inspect and run the model today. None of that is possible yet.
What Mistral itself says
Mistral describes the model as open-weight, not open-source, and says “we will release the weights by the end of the month.” Until then, it is red-teaming the model with “cybersecurity leaders, vetted partners, and state authorities”. Several outlets reported 27 October as the release date.
What the licence will be is unknown
The licence text has not been published. VentureBeat reported a custom Mistral licence rather than the Apache 2.0 licence used for Mistral Large 3, while The Decoder said the licence will arrive with the weights. A custom licence could limit commercial use, which would matter more than any benchmark.
What Artificial Analysis records today
Artificial Analysis still lists the preview as not open-weight, with no licence named. When we read its model data on 7 October, the record for Mistral Large 4 showed open weights as false and the licence as empty. A reader who trusts the Mistral Large 4 card over that record would be ahead of the facts.
The Free Tag on the Mistral Large 4 Card
The card’s structured data says the model is free, with a price of zero, and marks it “accessible for free”. The preview is not free to use through Mistral’s API.
What the preview costs
Mistral’s announcement lists $1.36 per million input tokens and $4.18 per million output tokens. OpenRouter lists the preview at half that, $0.68 and $2.09, which looks like a launch discount. Artificial Analysis measured about $1.13 per task across its index, because the model writes long answers.
Where “free” may come from
Open weights can be downloaded at no charge once released, and Mistral runs a consumer chat app with a free tier. Neither makes the model free today. Running a trillion-parameter model yourself also needs serious hardware, as our launch report worked out.
Bars are scaled to GPT-6 Astra’s $50, so each width is the price divided by 50: 20 divided by 50 is 40%, 14 divided by 50 is 28% and 4.18 divided by 50 is 8.36%. Prices are the same listings we used in our launch report on 6 October. Cheap is not the same as free.
The Rank Badge on the Mistral Large 4 Card
The badge reads “#27 in LLM models” and links to AIxploria’s LLM category. We walked the first three pages of that category, 36 entries, and checked the badges of the cards around position 27.
Position 27 is taken
Slot 27 holds Claude Sonnet 5, and Claude Sonnet 5’s own card is also badged “#27 in LLM models”. Mistral Large 4 does not appear anywhere in the first 36 positions. The category has 19 pages in total.
The badges around it are accurate
To rule out a broken counter, we checked seven control cards. Every one matched its position exactly, so the badge system works and the Mistral Large 4 card is the outlier.
| Position in category | Card in that position | That card’s own badge | Match |
|---|---|---|---|
| 1 | Claude Opus 5.5 | #1 | Yes |
| 11 | Kimi K3 | #11 | Yes |
| 25 | Grok 4 | #25 | Yes |
| 26 | GLM-5.2 | #26 | Yes |
| 27 | Claude Sonnet 5 | #27 | Yes |
| 28 | Gemini 3.6 Flash | #28 | Yes |
| 29 | Jev | #29 | Yes |
Two cards, one number
So two cards now claim the same ordinal. We have seen this shape before in this series, when a new card shared a slot already held by an older model. A reader who clicks the badge expecting to find Mistral Large 4 at 27 will find a Claude model instead.
The badge record across our checks
The Mistral Large 4 card joins a run of new cards in this series whose badges failed a control check. The table lists the main cases we documented, with the date we checked and the shape of the error. The shapes vary, but none of these badges led a reader to the right place.
| Checked | Card | Badge | What the slot actually held |
|---|---|---|---|
| 6 Sep 2026 | GPT-6 Astra | #1 in LLM models | Card absent from the first two pages |
| 6 Sep 2026 | Lyria 3.5 | #2 in Music | Its predecessor, Lyria 3 |
| 6 Sep 2026 | Weather Next 3 | #41 | Ranked below its own predecessor at #27 |
| 14 Sep 2026 | MAI-Transcribe-2 | #25 in Transcriber | Whisper Large V3 Turbo, also badged #25 |
| 24 Sep 2026 | Grok 4.7 and Claude Opus 5.5 | #4 in LLM models | Grok 4.6, so three cards shared #4 |
| 28 Sep 2026 | MiMo-V2.6 | #28 in LLM models | An offset slot also claimed by Grok 4.3 |
| 7 Oct 2026 | Mistral Large 4 | #27 in LLM models | Claude Sonnet 5, also badged #27 |
The pattern is consistent: older cards carry accurate badges, while a new card’s badge is a claim to check rather than a position you can rely on. That is worth knowing before anyone quotes a directory rank in a report.
What the badge does not measure
Even a correct badge would only show a directory’s ordering, not test results. On Artificial Analysis’s index Mistral Large 4 scored 38.4, the best result for any model built outside the US and China. That is the number our ranking explainer unpacks.
The Alternatives Strip on the Mistral Large 4 Card
Under the blurb, the Mistral Large 4 card shows three “similar AI tools“: Claude Opus 5.5, GPT-6 Astra and Claude Fable 5.1. Further down, a longer strip adds Claude Opus 4.8, Claude Sonnet 5.5, GPT-5.6, GPT-6.1 Sol, Grok 4.6, Grok 4.7 and Kimi K3.
Almost all closed models
Of the ten models across both strips, only Kimi K3 is open-weight. The strongest open rivals on Artificial Analysis, such as Xiaomi’s MiMo-V2.6-Pro and Z.ai’s GLM-5.3, are missing. For a model sold on open weights and sovereignty, that is an odd set of alternatives.
The irony in the first two names
Mistral’s own post singles out Claude Opus 5.5 and GPT-6 Astra. It says both “score near zero” on the vulnerability-reproduction test because they refuse the task, and it presents that as the reason to choose Mistral Large 4. The card lists those two as the first alternatives.
How the suggestions compare on the index
Artificial Analysis scores give a rough sense of the gap. The closed models lead, which is why Mistral argues on price, openness and refusals rather than on raw score.
| Model | Weights | Intelligence Index | Output price per million |
|---|---|---|---|
| Claude Opus 5.5 | Closed | 57.6 | $20.00 |
| Claude Fable 5.1 | Closed | 53.4 | Not compared |
| GPT-6 Astra | Closed | 52.7 | $50.00 |
| Kimi K3 | Open | 43.6 | $14.00 |
| Mistral Large 4 preview | Due end of October | 38.4 | $4.18 list |
Index scores are as reported by Artificial Analysis between 24 September and 6 October; Claude Fable 5.1’s comes from the 24 September leaderboard. Prices are list prices per million output tokens.
Where the Mistral Large 4 Card Has Spread
A card can be live for hours before it appears in the directory’s ranked lists. We searched AIxploria’s three server-rendered list pages for the model’s name at 10:33 UTC, about five hours after publication.
Not yet in the ranked lists
“Mistral Large 4” appeared zero times on the Top 100 page and zero times on the Ultimate List. It appeared three times on the Latest AI category page, which is where new entries land first. Mistral Large 3 does not appear on the Top 100 or the Ultimate List either.
What that means for a reader
In earlier checks, new cards took at least a week to reach the ranked lists, and some never reached the ranked body at all. So a reader browsing AIxploria’s top lists today would not find Mistral Large 4. They would only meet the Mistral Large 4 card by searching for it or by browsing the newest entries.
What the Mistral Large 4 Card Leaves Out for Businesses
The card is written for curious readers, not for buyers. Several facts that matter to an organisation are missing from the Mistral Large 4 card entirely.
Where the data goes
Mistral says the model will be available in several regions, including “a European deployment that Mistral operates end-to-end, independently of other digital service providers and under European law.” For UK and EU organisations with data-location rules, that matters more than any benchmark, and the card does not mention it.
Long answers raise the bill
Artificial Analysis found the model wrote about 200 million output tokens across its index, against a median of 81 million for comparable models. Per-token prices therefore understate the cost per task. Our LLM API pricing guide explains how to price the finished task instead.
The preview will change
Mistral says reinforcement learning is still running and the model “continues to improve rapidly”. Scores, behaviour and possibly prices may shift before the weights ship. A card frozen on 7 October will not show those changes unless someone edits it.
Cyber skills cut both ways
Mistral pitches fewer refusals on security work as a feature for defenders. The same skill can help attackers. Mistral reports a higher refusal rate on malicious cyber prompts than other open models, but any organisation should still set its own policy on who may use the model for security tasks.
How to Use a Directory Card Like This One
Directory cards are useful for discovery. They are a poor source for decisions. A five-minute routine turns the Mistral Large 4 card, or any similar entry, into a reliable starting point.
Follow the official link first
The card’s “official website” button leads to Mistral’s announcement, which is the right primary source. Read it before the blurb. Where the two differ, as on licence and price here, trust the vendor’s own page and an independent benchmark over the summary.
Check the three labels buyers copy
Licence, price and rank are the labels most likely to end up in a procurement note. Check each one against its source. On this card, two of the three were wrong and one was premature, while the technical description was largely right.
Look for what is missing
Ask what a card leaves out: data location, total cost per task, release dates and licence terms. For open-weight models, our guide to open-weight AI models lists the questions to ask before you build on one.
Mistral Large 4 Card: Frequently Asked Questions
Is Mistral Large 4 open source, as the card says?
Not yet. Mistral calls it open-weight and says the weights will ship by the end of October 2026. The licence has not been published, and Artificial Analysis still lists the preview as not open-weight.
Is Mistral Large 4 free?
No. The preview API costs $1.36 per million input tokens and $4.18 per million output tokens at list price. The Mistral Large 4 card’s “Free” tag is wrong for the model as it is offered today.
Is the “#27 in LLM models” badge correct?
No. Position 27 in AIxploria’s LLM category belongs to Claude Sonnet 5, whose own card also says #27. Mistral Large 4 was not in the first 36 positions when we checked.
What does the card get right?
The technical summary: 1 trillion parameters, text and image input, technical drawings and satellite imagery, vulnerability fixing and more than 160 languages all match Mistral’s announcement.
Should I rely on the card’s alternatives?
Use them with care. Nine of the ten suggestions are closed models, and the strongest open-weight rivals are missing. Compare open models directly if openness is why you are looking.
References
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
