Mistral is a dual-track family for commercial fine-tuning. The Mistral 3 announcement puts Ministral 3B, 8B and 14B plus Mistral Large 3 (41B active / 675B total) under Apache 2.0. Mistral’s help centre, dated August 2026, says certain other open cards use a modified MIT that adds a $20 million monthly-revenue commercial clause, with Studio as the paid path. Hosted platform SKUs are a third track. Open weights on one card are not a licence for every Mistral product. Read the identifier before anyone trains.
By Shen Pandi, Towards AGI. Reviewed September 2026. This page answers: Mistral open weights licence and fine-tuning.
Open weights · Apache-2.0 on Mistral 3 cards; other cards differ — confirm · Ministral dense 3B–14B through Large 3 MoE
Who this page is for
Read this when the brand on the slide is Mistral and nobody has written which track you are buying: Apache Mistral 3 weights, a modified-MIT card with a monthly-revenue gate, or a hosted Studio / La Plateforme SKU. The usual readers are counsel who were told the whole lab is Apache, a finance partner who has never seen the $20 million sentence, and an SRE who must know whether they will operate 14B dense or 675B total. This URL is not a Mixtral serving guide and it will not score Ministral against Qwen3.
If you already know you need Meta’s community pack, leave for Llama. If you already know you need an Alibaba Apache LICENSE on a named Qwen3 revision, leave for Qwen. If you already know you need a reproduced stack, leave for OLMo. Stay here while the identifier, the December 2025 announcement, the August 2026 help-centre note, and the hosted-versus-downloadable sentence are still fuzzy.
Three tracks hide under one brand
People say “Mistral is Apache” the way they used to say a whole lab was open source. The December 2025 Mistral 3 post is the clean commercial story: the new dense 3B, 8B and 14B models and Mistral Large 3 are released under Apache 2.0. That is a real announcement. It is not a grant for every weight Mistral has ever published, and it is not a grant for a hosted SKU on La Plateforme.
The help centre is explicit that this is not universal. Most open models are Apache 2.0. Certain models use a modified MIT licence under which companies above $20 million monthly revenue must obtain a commercial licence or use Mistral Studio. Older research or non-production licences also exist. Terms are documented on each model card. The card in the statement of work is the licence. The family name is a filing label.
This page owns that split. It does not rank Ministral against Mixtral on a public exam, and it does not teach a serving recipe for 675B total parameters. If the constraint is Meta’s community terms, read /models/llama. If the constraint is a Qwen3 Apache file versus an earlier Alibaba generation, read Qwen. If the constraint is a reproduced stack, read OLMo.
Apache on Mistral 3 is a dated announcement
The cited Mistral 3 post names the dense trio and Large 3, states Apache 2.0, and publishes the active-versus-total split for the MoE: 41B active, 675B total. That is the pack you want in a brief when someone says “we will fine-tune Mistral 3”. Write the post URL, the card URL, and the date you read them. If Mistral later ships a new dense size under a different file, you still have the announcement you relied on.
Apache-2.0 on those weights is notice-and-disclaimer for a closed product in the ordinary case. It is not OSAID. Downloadable Mistral 3 parameters let you host and adapt the model. They do not publish the preferred form for modifying the system in the Open Source Initiative’s sense. OSI’s open-weights note is the other citation here: parameters are not the whole artefact. If procurement meant a reproduced stack, Apache on Ministral is the wrong exhibit.
Do not copy a community naming prefix or a “Built with” line onto an Apache Mistral 3 card. Those duties live on another family’s page. Keep the Apache notice. That is the hygiene. If counsel already approved a different bargain and is now “just swapping to Mistral”, make them read the new file. Swapping families is a new legal read.
The modified MIT is a revenue gate, not a typo
The help centre’s modified MIT is the second track. It is still an open-weight story: you can often download and run the model. It adds a commercial clause keyed to monthly revenue. Companies above $20 million monthly revenue must obtain a commercial licence or use Mistral Studio. That is a different conversation from Apache notice. It is also a different conversation from Meta’s monthly-active-user threshold. Do not mash the two clauses together in a slide.
We will not calculate your revenue. Finance and counsel do that. What we will do is refuse to train on a modified-MIT card until someone has written which side of the line the licensee sits on, and whether Studio is the intended path instead of take-home weights. A specialist adapter for one letter family does not delete the clause. Internal-only serving does not automatically delete it. The help centre says to read the individual card.
Older research or non-production licences are a third flavour of “open Mistral” that still is not Mistral 3 Apache. A card that says evaluation-only, research-only, or non-production is a stop. Pilots that “we will commercialise later” are how those cards leak into a product brief. Stop at the identifier.
A hosted SKU is not the downloadable base
Mistral sells hosted and commercial products that are not the weight file. A La Plateforme SKU can be the right first step for quality: you send prompts, you get completions, you do not operate the box. It is the wrong artefact if the requirement is weights you can take into your environment, adapters you can hand to a later operator, or a licence file you can attach to a merged image.
Studio, as named in the help centre, is the commercial path for certain modified-MIT cards once the revenue clause bites. That is a contract. It is not Apache-2.0 on a Ministral 14B file. Write “self-hosted Mistral 3 weights” or “Studio / platform SKU” in the first line of the brief. If those words are missing, you do not yet have a Mistral project. You have a brand.
Fine-tuning on a vendor UI is a different path from downloading Apache weights and training adapters in your cluster. The closed-API comparison owns the prompt-path question. This page owns the mistake of filing a hosted completion as an open-weight fine-tune.
Self-hosted Mistral 3 weights under the December 2025 Apache announcement.
Another open identifier whose help-centre note may add the $20 million monthly-revenue gate.
Studio or La Plateforme SKU — a contract and an endpoint, not a take-home file.
Non-production or research-only text — a stop until the identifier changes.
Mixture-of-experts is an inference plan
Mistral Large 3 is published as 41B active and 675B total parameters. Mixtral-class models likewise activate a subset of experts per token. That changes VRAM, routing, batching and operator skill. It does not change the need to read the licence, and it does not make the model sovereign once adapted. A team that can serve 14B dense may not be able to serve 675B total. Write the serving plan before the training plan.
If the deployment target is a small private box, a Ministral 3B or 8B Apache card may be the candidate, not Large 3. That is a hosting decision. Measure it on the same reserved task. Do not pick Large 3 because the announcement was louder. Do not pick 3B because a slide said small models are cheaper. The commercial SLM page is where “should we shrink at all?” lives. This page is where “which Mistral grant did we think we had?” lives.
Cited Mistral bands. Other cards exist. Read the identifier on the card you will train.
Band
Cited example
Class
Licence on the cited source
Dense small
Ministral 3 — 3B, 8B, 14B
SLM to small LLM
Apache 2.0 (Mistral 3 announcement)
Large MoE
Mistral Large 3 — 41B active / 675B total
LLM (MoE)
Apache 2.0 (Mistral 3 announcement)
Other open cards
Named on the individual model card
Check the card
Apache 2.0, or modified MIT with a $20M monthly-revenue clause (help centre)
Insurance: two cards, one letter job
Claims correspondence on Ministral 14B versus a revenue-clause card
insurance
A carrier wants draft letters a handler still sends. If the proposed base is Ministral 14B Instruct under the Mistral 3 Apache announcement, the licence conversation is notice-and-handover: keep the Apache text with any redistributed adapter, name the revision, write the review path. The brief is one letter family, unused test letters, and a written bar for tone and exception rate — not “fine-tune Mistral”.
If someone instead pulled a Mistral card that uses the modified MIT revenue clause, or a non-production licence, stop. Those are different commercial facts. If monthly revenue sits above the help-centre line, Studio or a separate commercial licence is the conversation, not a quiet LoRA. If the team only has a La Plateforme SKU, they do not yet have take-home weights. Do not book GPUs on the word “Mistral”.
Retail: in-region catalogue copy is a track choice
Hosted SKU versus downloadable Ministral for in-region copy
retail
A retailer wants catalogue attributes and product copy filled from supplier text, and the data rule is that those strings must not leave the region. A downloadable Ministral 8B or 14B weight under the Mistral 3 Apache announcement can enter if the GPU environment is already inside that region: you take the file home, you train adapters there, you serve there. A La Plateforme or Studio SKU can still be the faster quality experiment. It is the wrong artefact for the in-region sentence, because prompts and completions then sit in someone else’s building.
Write the track in the statement of work before anyone pastes a demo. If the only available identifier is a modified-MIT weight and monthly revenue may sit above the August 2026 help-centre line, Studio is a commercial path — and it is still a hosted path. Do not file “we picked Mistral so the copy stays here”. File “Ministral 14B, Apache announcement, box in-region” or “Studio SKU, copy leaves, counsel accepted that”. Unused test products are the quality comparison. The track is the residency comparison.
When Mistral is the wrong first conversation
Pick another family when the organisation will not read more than one card and the proposed weight is not on the Mistral 3 announcement. Pick another family when counsel asked for OSAID and will not accept open weights plus Apache. Pick another family when the only available path is a hosted SKU and the requirement was take-home adapters. Those are track problems. They are not solved by a better expert-routing config.
Mistral is also the wrong first conversation when the team wanted Meta’s community bargain and is “just trying Mistral this quarter” without a new legal read. Swapping brands is not a loophole. Read /models/llama for that bargain, then come back only if the brief actually names a Mistral identifier.
Decision prompts, not scores.
When
Then
The named weight is Ministral 3 or Large 3 under the Apache announcement
The licence conversation is notice-and-handover. Write the post, the card and the serving plan.
The named card uses the modified MIT and monthly revenue may exceed $20 million
Stop. Finance and counsel write the line. Studio or a commercial licence may be the path, not a quiet adapter.
The brief only names a La Plateforme or Studio SKU
That is a hosted product. Do not file it as open-weight fine-tuning.
Procurement meant a reproduced training stack
Read OLMo. Apache on Mistral 3 weights is not that exhibit.
What changes in the run
On a Mistral 3 Apache card, a typical specialist run freezes the named dense or MoE base and trains adapters. The artefact you store is the adapter or a merged copy, the exact identifier, the Apache notice, and notes from reserved evaluation documents. The grant does not change because the optimiser ran. Serving 14B dense versus 675B total still changes the operator’s week.
On a modified-MIT card, the same technical run can be the wrong commercial act. If the revenue clause applies, the honest artefacts may be a Studio contract and API logs, not a weight file you take home. On a hosted SKU, there may be no adapter to hand over at all. Write the track before you write the rank. The LoRA page owns the adapter file. This page owns which Mistral grant you thought travelled with it.
In-region catalogue copy makes the track visible in the run itself. A Ministral adapter trained on a box you operate produces a file you can store next to the Apache notice. A Studio experiment produces logs, a prompt pack, and a contract clause. Those are not interchangeable handover objects. Do not merge them in a slide that says “we fine-tuned Mistral in-region” if the completions never left the vendor endpoint.
The announcement date and the help-centre date are both exhibits
The December 2025 Mistral 3 post is the exhibit for Ministral 3B, 8B and 14B and for Large 3’s 41B-active / 675B-total Apache story. The August 2026 help centre is the exhibit for the sentence that this is not universal, and for the $20 million monthly-revenue fork toward a commercial licence or Studio. Keep both URLs in the brief. If a later post changes a size or a file, you still have the dates you relied on.
Do not let a sales deck collapse those two dates into “Mistral is Apache now”. The post is about a named release. The help centre is about the rest of the catalogue. A team that only bookmarks the post will miss the revenue gate. A team that only bookmarks the help centre will miss which dense and MoE weights were actually announced under Apache 2.0. Both sentences belong in the same paragraph.
Handover that names the track
A useful Mistral handover states the track in the first line: Apache Mistral 3 weights, modified-MIT card plus revenue acknowledgement, or hosted SKU. Then it includes the identifier, the licence file or contract clause, adapters only if weights were actually trained, and evaluation notes from unused test documents. It is not a claim that “Mistral is Apache”, that inference is private, or that a public exam transferred to your letters.
Operators still need the base weights on the Apache track and a path that allows the stack. Merging simplifies a single-task image. Keeping adapters separate helps more than one product on one frozen Ministral base. Choose from the operating plan. If the track was Studio, the handover is credentials, data-handling terms and an exit sentence — not a GGUF.
Questions procurement should paste
Is this Ministral 3, Large 3, another open card, or a platform SKU? If it is Mistral 3, have we kept the announcement URL? If it is another open card, does the help-centre modified MIT apply? Where does monthly revenue sit relative to $20 million? Are we taking weights home or buying Studio? Did anyone mean OSAID when they said open source?
Those questions are more useful than a debate about whether Paris is “more open” than another lab. This page exists so the Apache announcement, the revenue-clause track and the hosted SKU do not collapse into one word on a slide.
Sources
Mistral AI, Introducing Mistral 3 (Accessed September 2026). Mistral 3 announces dense 3B, 8B and 14B Ministral models and Mistral Large 3 (41B active / 675B total) under Apache 2.0.
Mistral Help Center, open-model licences (Accessed August 2026). Mistral states most open models are Apache 2.0, while certain models use a modified MIT licence that requires companies above $20 million monthly revenue to obtain a commercial licence or use Mistral Studio. Terms are documented on each model card.
Open Source Initiative, Open Source AI Definition 1.0 (Accessed September 2026). An open-source AI system must provide the preferred form for modification, including parameters under OSI-approved terms plus the data information and code needed to derive them.
Open Source Initiative, Open Weights (Accessed September 2026). Open weights are the published parameters of a trained network. Weights alone are not the preferred form for modifying the system.
Questions we hear first
Are all Mistral models Apache-2.0?
No. Mistral 3 dense and Large 3 weights are announced as Apache 2.0. The help centre says certain other cards use a modified MIT licence with a $20 million monthly-revenue clause. Hosted products are separate. Read the card for the weights you will actually train.
Does Mixtral or Large 3 change what we may fine-tune?
Not by itself. Mixture-of-experts changes memory and serving. The right to adapt still comes from the licence on that checkpoint. Large 3’s 41B active / 675B total split is a hardware fact, not a quality result.
What happens if we sit above the $20 million monthly-revenue line?
On a card that uses the modified MIT, the help centre says you must obtain a commercial licence or use Mistral Studio. We will not decide which side of the line you sit on. Finance and counsel write that sentence.
Is a La Plateforme completion an open-weight fine-tune?
No. A hosted SKU is a product. You may still evaluate quality there. You do not automatically receive Apache weights, adapters, or a take-home grant.
Can we distill another family from Mistral 3 outputs?
Apache on a Mistral 3 teacher is a more ordinary starting point than a community clause that forbids improving a different large language model. You still need a licensed teacher, a quality bar on that teacher, and a read of the student card. Ask counsel before traces leave the building.
Should we default to Ministral 3B because it is small?
It is a candidate when latency or footprint matters. Defaulting to the smallest instruct model without a reserved comparison is how pilots waste a quarter. The commercial SLM page is the shrink conversation.
Does Apache on Mistral 3 make the system open-source AI?
Apache-2.0 is an OSI-approved software licence on the released artefacts. OSAID also asks for the preferred form to modify the system. Mistral 3 weights are usually open weights with a permissive licence, not a reproduced stack. Read OLMo if that was the ask.
If product copy must stay in-region, is any Mistral SKU enough?
No. A downloadable Ministral weight helps only when the GPU environment is already inside the region. A Studio or La Plateforme SKU is still someone else’s building. Write the track before you write the model name.
Say whether you mean open Mistral weights or a hosted SKU.
Those are different licences and different inference paths. Bring the card and the environment. We will decide if adaptation or a readiness review is first.