In under three years Mistral AI went from a three-person idea to a €11.7B (~$14B) valuation[6], Europe's first tech "décacorne"[11], with an annualised revenue run-rate estimated to have climbed roughly 20x to ~$400M and a stated goal of topping €1B in 2026[18][60].

The genuinely open question is not whether Mistral is a serious lab — it plainly is, with frontier-adjacent models, marquee customers and a national-champion role in France. It is whether a company funded an order of magnitude below its US rivals, squeezed between American frontier labs above and ultra-cheap Chinese open-weight models below, can stay at the frontier — and whether its "sovereign European" identity survives contact with its US and Dutch cap table. The evidence does not cut evenly: weighed, it points to a durable European enterprise franchise rather than a frontier winner — a lean each section shows its work for, with the tripwires that would overturn it collected in the closing weighing of Risks & Skeptics.

The decisive questions

Each links to the section that lays out the evidence on both sides.

Revenue run-rate, the headline number

Estimated ARR, US$M. Private-company figures are third-party estimates (Sacra); the 2026 endpoint is the company's stated revenue target, not booked ARR. The shape — explosive growth off a tiny base — is the story.

How to read this

Nine sections, each built the same way: a neutral synthesis, a two-sided case-for / case-against ledger, sourced data and charts, and dated facts. Because Mistral is a French company, roughly a third of the sources are in French — the domestic political and sovereignty debate is far richer in-language, and translated quotes show the original text. Start with the question that interests you, or read in order from the Overview.

Mistral was founded in April 2023 by Arthur Mensch (ex-DeepMind) and two ex-Meta LLaMA authors, and within ~28 months had raised over €2.8B across four rounds to a €11.7B valuation — becoming France's first tech "décacorne"[6][11][12]. That speed is the bull case (extraordinary execution and pull) and the bear case (a frontier-scale valuation reached before the revenue or the moat is settled) in one fact.

What Mistral is

Mistral AI is a frontier-AI lab building large language models and the products around them: an API platform (La Plateforme), a consumer/enterprise assistant (Le Chat, rebranded Vibe in May 2026), coding and reasoning models, and a nascent sovereign GPU cloud (Mistral Compute)[20][36]. Its founders are Arthur Mensch (CEO, who worked on DeepMind's Flamingo and Chinchilla), Guillaume Lample (Chief Scientist) and Timothée Lacroix (CTO), the latter two senior authors on Meta's LLaMA paper[1]. From the start it positioned itself as "a European champion with a global vocation" in generative AI[3].

The timeline

Funding and valuation by round

Post-money valuation, €B. Each round roughly 2-3x'd the last; the September 2025 ASML-led round more than doubled the prior €5.8B mark to €11.7B[5][6].

Leadership and footprint

The three founders still run the company (Mensch CEO, Lample Chief Scientist, Lacroix CTO), with ASML's CFO taking an advisory seat on Mistral's Strategic Committee after the 2025 round, making the Dutch firm Mistral's principal outside shareholder[6][8]. Headcount grew from a few dozen in early 2024 to several hundred by 2025; Mistral has expanded beyond Paris with offices across Europe and the US, and added an Austrian engineering-AI team via its 2026 Emmi AI acquisition[10]. By early 2026 it had made two acquisitions — Koyeb (cloud infrastructure) and Emmi AI (industrial simulation) — signalling a push beyond pure model-building into an "AI cloud" and vertical AI[9][10].

Estimates put generative AI on a path to hundreds of billions in annual revenue, growing ~35-43% a year[14]. But by one 2026 tally OpenAI and Anthropic alone hold >90% of the category's valuation[15], and the open-weight niche Mistral built is now led by Chinese labs that out-download and undercut it. The market is a tailwind; Mistral's share of it is the question.

A large, fast-growing market

Sizing the market depends heavily on scope. Narrowly — foundation-model APIs and licensing — it is on the order of tens of billions of dollars today; broadly defined (all generative-AI software and services) estimates run well into the hundreds of billions, with compound growth rates commonly cited in the 35-43% range and trillion-dollar long-run projections[14]. Either way the prize is big enough to justify the capital flooding in — which is precisely why the competition is so punishing.

Concentration at the top

Value has pooled at the frontier. One 2026 industry ranking found that OpenAI and Anthropic together account for more than 90% of the total valuation in the foundation-model category[15] — leaving Mistral, xAI, Cohere, DeepSeek and the rest to share what's left. Mistral is Europe's most valuable AI company, but in the global field its ~$14B valuation is a rounding error against the leaders (see Peer Comparison).

The open-weight layer got commoditised

Mistral was the earliest leader of the open-weight ecosystem, peaking at roughly 57% of derivative models in December 2023 on the strength of Mistral 7B and Mixtral[16]. But its cumulative downloads never kept pace with Meta's Llama or Alibaba's Qwen, and in January 2026 DeepSeek overtook Mistral in cumulative downloads — costing it its place as the third-largest open model family[16]. Chinese labs have also driven prices down hard: DeepSeek's V4 models were priced around $0.28-$3.48 per million output tokens versus roughly $25-30 for OpenAI and Anthropic — a near-10x undercut on the open layer Mistral helped create[17].

Estimated ARR rose from ~$16M (end-2024) to ~$400M (early 2026), roughly 20x in a year[18][19], with the company targeting >€1B in 2026[60]. The revenue mix leans on Europe (~60%)[23] and on a few large contracts — strong momentum, but concentration and unproven margins are the flip side.

How Mistral makes money

Per analyst breakdowns, revenue comes from three recurring streams plus bespoke deals[18]:

• La Plateforme (API): pay-per-token access to Mistral's models, positioned as among the cheaper frontier-adjacent options.
• Le Chat / Vibe subscriptions: a consumer/prosumer assistant with a $14.99/month Pro tier and enterprise plans[20].
• Enterprise & on-prem deployments: private and self-hosted installations tied to data-residency and regulatory needs — the sovereignty selling point.
• Large co-development contracts: nine-figure, multi-year partnerships that expand total contract value[18].

The marquee deals

Mistral has converted its national-champion status into anchor customers. Shipping group CMA CGM signed a €100M, five-year partnership in April 2025, embedding teams in Marseille[25]. In February 2026 it signed a multiyear deal with Accenture, which also became a customer — the same marquee integrator OpenAI and Anthropic had recently signed[21]. French coverage describes enterprise work spanning Airbus (R&D modelling), BMW (crash-test simulation) and BNP Paribas (KYC)[26], and the French armed forces now run a sovereign Mistral platform recording 19 million queries a year[22][24].

Owning the stack

Unusually for a model lab, Mistral is integrating downward into compute. It is investing several billion euros in its first French data centre on the Saclay plateau (Essonne), aiming to control its value chain "from machine to software" and exploit France's low-carbon grid[27], has committed to a ~€1.2B facility in Sweden[23], and took on $830M of debt in March 2026 for ~13,800 Nvidia chips near Paris[19]. The logic is sovereignty and margin control; the risk is front-loading enormous fixed costs.

The unproven part: economics

Mistral discloses no audited margins or cash burn. French coverage frames the €1B-revenue-in-2026 target as a "risky bet" against likely sub-€150M revenue in 2025, paired with capex on the order of $1B[28] — i.e. infrastructure spend that could roughly match revenue. Whether the API/subscription mix plus a few big contracts can carry that cost base, at gross margins squeezed by ~10x-cheaper open-weight rivals, leans negative for now on the estimates available: capex of the order of revenue[28] and the choice to fund chips with $830M of bank debt rather than cash[19] both indicate a model that does not yet pay for itself — medium confidence, since the true margins are undisclosed (the arithmetic is in Financials).

On neutral capability leaderboards Mistral's models rank below the Claude / GPT / Gemini frontier[30], and by consumer traffic it sits around #7 among AI chatbots[29]. Its edge is not peak capability but efficiency, openness, multilingual strength and European sovereignty — a genuine position, but one now contested on every side.

The five forces

Click a force to see the rated pressure and the evidence behind it. Read from Mistral's seat, four of the five forces are high — an unusually hostile structure.

Where Mistral sits

Positioning on two axes that matter in this market: capital/scale (horizontal) and model openness (vertical). Hover a point for the basis. Mistral is a hybrid — more open than the US frontier labs, far smaller in capital — sharing the open quadrant with Chinese labs that now out-download it.

The competitors

• US frontier labs (OpenAI, Anthropic, Google): larger models, vastly more capital, and — between OpenAI and Anthropic — >90% of category valuation[15]. Mistral cannot match them on training spend.
• Chinese open-weight labs (DeepSeek, Qwen): cheaper and, on downloads, now ahead of Mistral on its own open turf[16][17].
• Cohere + Aleph Alpha: a 2026 merger into a ~$20B entity explicitly positioned as a sovereign alternative to US Big Tech — a direct challenger to Mistral's European niche[31].

Mistral runs a two-tier strategy: Apache-2.0 open-weight models (7B, Mixtral, Large 3) that build mindshare, plus proprietary "premier" models (Medium, Magistral Medium) that it bills for[33]. The open tier is the marketing and the moat-by-adoption; the closed tier is the revenue — and the tension between them is the strategic story.

The model lineup

Mistral ships fast across the size spectrum: open models like Mistral 7B and the Mixtral mixture-of-experts that made its name, the Magistral reasoning line (June 2025), the large open-weight Mistral Large 3 (Dec 2025), and proprietary Medium models tuned for cost-efficiency[33][35]. Magistral Medium scored 73.6% on AIME 2024 (up to ~90% with majority voting), with transparent multilingual chain-of-thought[35]. Medium 3.5 (April 2026) folds chat, reasoning and coding into one model at 77.6% on SWE-Bench Verified and roughly half a leading rival's price[39].

The efficiency-and-sovereignty wedge

Mistral's durable advantages are not raw capability but: cost-efficiency (sparse MoE designs, cheap API pricing), multilingual strength (especially European languages), open weights for customers who need to self-host, and data-sovereign / on-prem deployment that US providers serve less easily. Those are real switching-cost and trust moats in regulated European sectors — and weak moats against a buyer who just wants the best model on a US cloud.

The agent pivot: Le Chat → Vibe

In May 2026 Mistral rebranded Le Chat as Vibe, repositioning the assistant from a chatbot into an agentic work product with Work (enterprise connectors), Code (sandboxed agents that open pull requests) and Chat modes[36]. French coverage read it as Mistral "going on the offensive" globally rather than just playing in France[38] — though some lamented the loss of the brand's French charm[50].

The open-source tension

The strategy's sharpest critique is that Mistral has quietly walked back the openness that built it: Mistral Large, released the same day as the Microsoft deal, was its first proprietary, paid-API model with no code access[57], and critics describe withholding training data and using restrictive premier-model licenses as "open-washing"[67]. Mistral's reply is that a sustainable lab needs a revenue tier, and that it still ships genuinely open models others can build on[33] (more in Sovereignty).

Mistral is genuinely woven into French strategic ambition — a Macron-backed champion, an arm of defence and banking digitisation, and the headline of a €109B national AI push[45]. Yet its largest outside shareholder is Dutch (ASML, ~11%)[6], Microsoft and US funds hold stakes, and it lobbied to soften the EU AI Act[41]. Whether that is pragmatic champion-building or a hollow "sovereignty" label is the debate.

The national-champion case

France treats Mistral as strategic infrastructure. President Macron urged citizens to use Le Chat over ChatGPT, and at the February 2025 Paris AI Summit announced €109B of private AI investment, explicitly framed as France's answer to the US "Stargate"[45]. The Assemblée Nationale has debated "preserving Mistral" as a matter of national sovereignty, with the government calling digital sovereignty a priority[51]. CEO Arthur Mensch makes the economic-sovereignty argument directly:

The AI Act lobbying controversy

In 2023-24, as the EU finalised its AI Act, Mistral (with Germany's Aleph Alpha) led a startup push to keep binding obligations off foundation-model makers and on downstream deployers[53]. A civil-society investigation argued Mistral's arguments were "remarkably similar to Big Tech's," and noted that once the text was softened, Mensch called it manageable[41] — an episode a separate watchdog framed as European startups "teaming up with Big Tech to gut the AI Act"[52]. The Ada Lovelace Institute countered that exempting foundation-model providers would be "irresponsible"[53], and EU Commissioner Thierry Breton accused Mistral of defending its business rather than the public interest[44].

The episode also drew conflict-of-interest scrutiny around Cédric O, France's former digital minister who helped shape early discussions of the law and then became a Mistral adviser/shareholder; a rapporteur recounted him saying "we shouldn't have rules that are too strict"[42], and a senator flagged unease that he "defends the interests of extra-European actors"[43]. O rejected the "lobbyist" label and argued he was fighting for European startups against marginalisation[54].

The Microsoft moment

Days after the AI Act fight, Mistral announced a Microsoft partnership (a €15M investment plus Azure distribution) and released its first proprietary model[4]. The Open Markets Institute said the deal "exposes as a farce" Mistral's "European champion" lobbying[55]; Green MEPs asked the Commission to examine its ethics[56]; and Sifted noted the simultaneous retreat from open source[57]. Mistral called itself "an independent European company with global ambitions"[56].

Defence — and where Mistral won't comment

Mistral has leaned into defence where US rivals have been cautious: a February 2025 partnership with Germany's Helsing on vision-language-action models, and a 2026 framework with the French armed forces (its "GenIAl" platform logs 19M queries a year)[22][24]. Critics call this a "ratchet effect" normalising military AI[48], and note Mistral declined to say whether autonomous-weapons or mass-surveillance uses are excluded from these deals[58]. Mensch defends the work, rebutting even Pope Leo XIV's call to keep AI out of warfare[59].

There is also a domestic regulatory edge: a February 2025 complaint alleged Mistral let only paying Le Chat users opt out of data exploitation — an unequal access to the GDPR right to object; Mistral said it took the matter seriously and updated its policy[49].

ARR is estimated to have grown ~20x to ~$400M in a year, with a >€1B 2026 target[18][60] — extraordinary growth. But on a €11.7B (~$14B) valuation that is a ~30x-plus ARR multiple with no disclosed path to profit and ~$1B of planned capex that could roughly match revenue[61].

Revenue trajectory (estimates)

Estimated ARR, US$M — private-company figures from Sacra; the 2026 endpoint is the company's stated revenue target, not booked ARR. The slope is real; the absolute level is small relative to the valuation.

Sources: Sacra ARR estimates; CEO revenue target.[18][19][60]

The capital stack

Mistral has raised over €2.8B in equity across four rounds to a €11.7B valuation (Sept 2025, ASML-led)[6][7], and in March 2026 added $830M of debt from a seven-bank consortium to buy ~13,800 Nvidia chips for a data centre near Paris[19]. It is also committing several billion euros to French and Swedish data centres[27][23]. This is the deepest war chest in European AI — and still, by frontier standards, modest:

The funding gap, quantified

The crux of the whole study in one picture. Mistral's figure is its estimated lifetime capital raised across all rounds; the US labs' figures are their single most recent mega-round alone — so even on this conservative comparison each one out-raised Mistral's entire history many times over. Log scale.

Mistral ~$3-4B lifetime (midpoint $3.5B shown)[13]; OpenAI $122B round[63]; Anthropic $65B Series H[62]; xAI $20B round[64]. US-lab bars are one round each, not lifetime totals, so the true gap is larger — this is the "order-of-magnitude below" / ~10x funding disadvantage the thesis turns on.

The economics we can't see

As a private company Mistral discloses no audited revenue, gross margin, operating loss or cash burn. Third-party estimates put 2026 capex on the order of $1B against the >€1B revenue target — i.e. infrastructure spend that may roughly equal revenue, with gross margins pressured by ~10x-cheaper open-weight competition[61]. French coverage frames the €1B target as a "risky bet" off a likely sub-€150M 2025 base[28]. Every profitability figure here is an estimate; treat them accordingly.

At ~$14B, Mistral is worth roughly 1/60th of Anthropic (~$965B) and OpenAI (~$852B)[62][63]. On revenue the gap is just as stark — ~$400M ARR vs Anthropic's ~$47B — though it is broadly in line with xAI and Cohere[64][66]. Mistral's problem isn't being European; it's scale.

Valuation (US$B, private marks)

All private; point-in-time round/secondary marks on the dates shown.[62][63][64][65][6]

Revenue / ARR (US$B, estimates)

Estimates/run-rates; Anthropic discloses run-rate, others are third-party estimates.[62][64][18][66]

The benchmarking table

Private valuations are point-in-time round marks and not directly comparable to public-market caps; ARR figures are estimates except Anthropic's disclosed run-rate. As of June 2026.

The bear thesis is that Mistral is structurally sub-scale: out-spent by US labs, out-downloaded by Chinese labs, and dependent on a sovereignty narrative complicated by its own cap table[13][16]. The bull rebuttal is that efficiency, sovereignty and hyper-growth can sustain a profitable European niche even if Mistral never tops the global leaderboard.

The capital and frontier risk

Mistral's ~$3-4B lifetime funding is an order of magnitude below its US rivals; one analysis argues it is "cash-starved by frontier-AI standards" and "cannot afford many failed runs"[13]. If frontier capability keeps requiring ever-larger training budgets, Mistral risks falling progressively further behind. Rebuttal: efficiency, smaller models and a focus on deployable enterprise AI may matter more than raw leaderboard position for its actual customers (see Strategy).

Commoditisation risk

The open-weight layer Mistral pioneered is now led by Chinese labs that out-download and ~10x-undercut it[16][17], and critics argue its own partial retreat into proprietary models — while withholding training data — amounts to "open-washing"[67]. Rebuttal: a revenue tier is necessary to fund a lab, and open models still seed adoption.

Independence and structural risk

Some analysts argue Europe's recurring pattern is that promising startups get absorbed or out-resourced by US tech[68], and Mistral's heavy US/Dutch ownership and Microsoft ties feed that worry (see Sovereignty). Rebuttal: the founders still control the company, and strategic backers like ASML are framed as a defence against US acquisition[6].

Talent and key-person risk

The AI talent war — with reported nine-figure packages from larger labs — raises retention risk for smaller players, and Euronews flags that Mistral's capacity "to keep hold of their key staff might be limited"[69]. The company is also concentrated in three founders. Rebuttal: Mistral's mission, equity and Paris research base have so far attracted and retained strong talent.

The weighing

Listing risks and rebuttals is not the same as weighing them. Here is where the evidence actually lands on each of this study's four decisive questions — with a confidence level, the strongest surviving counter, and the concrete tripwires that would flip the reading.

On whether a capital-light lab can stay at the frontier: the evidence leans no — Mistral competes near the frontier, not at it (medium confidence). The controlling evidence is Mistral Large 3 scoring ~23 on the Artificial Analysis Intelligence Index, average-to-below-average among comparable models[34], and a ~$3-4B lifetime capital base against single rounds of $122B (OpenAI) and $65B (Anthropic)[13][63][62], which outweighs the efficiency counter because the benchmark gap has persisted across successive model generations[40] even as the spending gap widened. The strongest surviving counter-argument: Medium 3.5's 77.6% on SWE-Bench Verified at roughly half a leading rival's price[39] — for many enterprise jobs, price-performance beats peak capability. What would flip this reading: a Mistral model in the top five of the Artificial Analysis Intelligence Index within the next two release cycles; or end-2026 revenue clearing the stated >€1B target[60] despite the capability gap. Pre-mortem: if this looks wrong in two years, the most likely reason is that efficiency-tuned models plus owned compute[27] made leaderboard rank commercially irrelevant — or, on the other side, that the frontier kept compounding and Mistral's anchor deals migrated to more capable US models.

On "sovereign European AI": the evidence leans toward real-as-a-product, strained-as-an-ownership-story (medium confidence). The controlling evidence is the French armed forces' framework agreement and the sovereign GenIAl platform's 19 million queries a year[22][24], plus ~60% of revenue earned in Europe[23], which outweighs the cap-table objection because customers are buying on-prem, data-resident deployment — not the shareholder register. The strongest surviving counter-argument: the Open Markets Institute's charge that the Microsoft deal "exposes as a farce" the European-champion lobbying[55], with Dutch ASML (~11%) the largest outside shareholder[6]. What would flip this reading: a marquee sovereign contract — defence, banking or an EU institution — lost to the ~$20B Cohere-Aleph Alpha combination[31]; or any transaction giving a US acquirer control over the founders. Pre-mortem: the bull-side miss would be underestimating how much tightening EU data rules expand sovereign demand; the bear-side miss, assuming sovereign procurement stays loyal once it proves price-sensitive.

On the open-weight wedge: the evidence leans toward the wedge no longer differentiating (high confidence). The controlling evidence is DeepSeek overtaking Mistral in cumulative downloads in January 2026 — costing it the third-largest open-family slot[16] — and Chinese APIs priced at $0.28-$3.48 per million output tokens[17], which outweighs the heritage argument because the wedge's two payoffs, adoption and price leadership, now both sit with Chinese labs. The strongest surviving counter-argument: open weights still seed enterprise adoption, and Mistral keeps shipping genuinely open Apache-2.0 models[33][34]. What would flip this reading: Mistral regaining the #3 open-model-family position by downloads in the next ATOM-style ecosystem tally[16]; or premier-model revenue clearing €1B[60], making the wedge moot as a revenue question. Pre-mortem: the bull-side miss is that open-weight trust in regulated European sectors mattered more than raw downloads; the bear-side miss, that openness stopped being a moat the day it became table stakes.

On the valuation: contested — this one genuinely deadlocks, and the deadlocking evidence is specific. For: ARR up ~20x to ~$400M in roughly a year[18][19]. Against: capex estimated at the order of revenue[61] and a 2026 target French coverage calls a "risky bet" off a likely sub-€150M 2025 base[28]. At ~35x estimated ARR the mark sits between Anthropic's ~21x and xAI's ~460x[62][64] — neither obviously cheap nor uniquely exuberant (the arithmetic is in Financials). What would flip it to justified: end-2026 revenue at or above the stated €1B[60]. What would flip it to exuberant: 2026 revenue below ~€500M, or a flat-to-down next round. Pre-mortem: if the mark looks cheap in two years, the miss was how fast nine-figure enterprise and defence contracts compound[25][22]; if it looks rich, the miss was revenue concentration in a handful of contracts that did not renew[18].

Method

Research proceeded by fan-out web search across nine question areas (overview, market, business model, competition, strategy, sovereignty/regulation, financials, peers and risks), followed by direct fetching of primary and reputable secondary sources — the URLs cited here were opened and read during the research run, not merely linked. Each claim was then transcribed into a structured manifest tagging it with a source tier (Tier 1 primary, Tier 2 reputable secondary, Tier 3 tertiary), a confidence level, and a stance (supporting, critical or neutral). Because Mistral is a French company, a substantial native-language pass was run: roughly a third of the sources (23 of 69, 33%) are French-language, weighted toward the domestic political, regulatory and sovereignty debate, which is far richer in French than in English coverage. Translated quotes are shown with their original French text so the verifiable wording is never hidden.

Frameworks used

The analysis applies Porter's Five Forces to read the foundation-model industry from Mistral's seat, a capital-versus-openness positioning map to place it against US and Chinese labs, peer benchmarking on valuation and revenue, a SWOT, and a case-for / case-against ledger in every section so strengths and weaknesses get equal scrutiny. These frameworks organize evidence rather than deliver a verdict. A formal valuation was deliberately skipped: as a private company with undisclosed economics, Mistral's worth hinges on contested judgments — whether it can stay at the frontier, whether the sovereignty moat holds — that a single intrinsic-value number would obscure rather than clarify.

Disclosed vs. estimated

Mistral is privately held and publishes no audited financials, so nearly every figure here is an estimate or a company-stated target, labeled as such. Funding-round sizes and valuations (the €105M seed through the €11.7B Series C, and the $830M debt) are the firmest numbers, corroborated by participants' own statements such as ASML's press release. The ARR trajectory (~$16M → ~$400M) comes from third-party analysts (Sacra) and is a run-rate estimate, not booked revenue; the ">€1B in 2026" figure is the CEO's stated target, not a result. Capex, gross-margin and profitability figures are unverified third-party estimates. Peer valuations and ARR are point-in-time private marks and analyst estimates on the dates noted. Source mix: supporting 18 · critical 24 · neutral 27; 7 Tier 1 primary, 48 Tier 2 reputable secondary, 14 Tier 3 tertiary.

Neutrality & independence

Each section deliberately pairs the case for and the case against — the bull view (hyper-growth, a real sovereign/efficiency niche, Europe's flagship lab) and the bear view (sub-scale capital, commoditised open weights, a strained sovereignty story, a rich multiple) — and then weighs them: each decisive question carries a stated lean, a confidence level and concrete tripwires that would flip it, collected in the closing weighing of Risks & Skeptics. The weighing is this study's reading of the evidence, not advice. It is not investment advice and is not affiliated with, sponsored by, or endorsed by Mistral AI. It is a point-in-time artifact as of June 2026; Mistral's funding, products, models and metrics move fast, and figures will age.

Full bibliography with tiers, stance, language and links on the Sources page.

Cross-checked at build time by an automated link checker. A few primary pages (Mistral's own blog/model pages) bot-wall automated fetchers and were verified against participants' statements and reputable secondary coverage. See Methodology & Limits.