Mistral Large 4 Launches: Can Europe Reclaim Its AI Glory?

French AI company Mistral released a public preview of its next-generation flagship model, Mistral Large 4, on October 6, with model weights scheduled for open release by end of October. Once hailed as the "face of European AI," the company is returning after a full five months since its previous flagship, Mistral Medium 3.5 — an unusually long gap in an industry where monthly or even weekly releases have become the norm. So how does this model, expected to "reclaim Europe's AI glory," actually stack up?

Specs: Mistral's Largest Model Yet

Mistral Large 4 uses a MoE (Mixture of Experts) architecture, with approximately 1.05 trillion total parameters, 49B active parameters, and a context window of up to 1 million tokens. It's a natively multimodal model capable of processing both text and images, with native support for over 160 languages. Mistral claims it leads among open-source models in cybersecurity, finance, and manufacturing.

The trillion-parameter count certainly justifies the "Large" in its name — this is Mistral's biggest model to date. Fittingly, its internal codename is "Le Chonc," meaning roughly "the big chunky one." Given Mistral's orange cat logo, the name feels perfectly on-brand.

The official benchmark results look impressive: SWE Atlas Q&A at 61.7%, Terminal Bench at 28.3%, and a composite Coding Agent Index of 49.8%. Independent evaluator Artificial Analysis gives it a composite intelligence score of around 38, roughly on par with GPT-class models — though slightly below DeepSeek^Ad V4.1 Flash's score of 39.

MoE (Mixture of Experts) is one of the dominant architectural designs for large language models today. Unlike traditional "dense models" that activate all parameters during every inference pass, MoE splits the model into multiple "expert" sub-networks, with only a small subset of experts routed and activated per input. This means that while the model has 1.05 trillion total parameters, only around 49B are active during any given inference — dramatically reducing per-inference compute costs while retaining the knowledge capacity of a much larger model. Leading models like DeepSeek, GPT-4, and Mixtral all use this architecture. When evaluating MoE models, active parameters are a more meaningful measure of actual inference cost and speed than total parameters; 49B active parameters is roughly equivalent to the compute load of a mid-sized dense model.

PR Blunder: A ChatGPT-Generated Promo Image

In its eagerness to showcase progress, Mistral posted a comparison chart on X showing its model jumping from last place to near the top. Sharp-eyed users quickly spotted something off — the font for the "V" in "DeepSeek V4.1" looked wrong, and "Muse Glimmer" was misspelled.

The likely explanation: the promotional graphic appears to have been generated using ChatGPT, and Mistral neither carefully reviewed the output nor caught the errors before posting. For a company trying hard to prove its capabilities, this kind of careless slip is more than a little embarrassing.

Cybersecurity: The Real Focus This Time

Setting the PR mishap aside, Large 4 places significant emphasis on cybersecurity capabilities. On tasks like those in CyberGym E2E — covering vulnerability discovery, reproduction, and remediation — it stands out. Independent testing showed an 81.7% success rate on one end-to-end benchmark, the highest among 18 models tested at the time.

This matters in practice. OpenAI and Anthropic have become increasingly cautious about cybersecurity use cases, and their safety mechanisms now block legitimate security tasks more often than they complete them. Mistral has found a meaningful differentiator in a space where the leading players have effectively ceded ground through self-imposed restrictions.

OpenAI and Anthropic enforce strict usage policies and content filtering mechanisms around cybersecurity applications. These systems are designed to err heavily on the side of refusal, which means legitimate penetration testing, vulnerability reproduction, and security research tasks are frequently misclassified as malicious and blocked. Security practitioners refer to this as the "over-refusal" problem. For red team engineers and security researchers who need to simulate attacker perspectives to find and fix real vulnerabilities, a model that is both technically capable and less restrictive on security-related tasks has genuine commercial value. Mistral's deliberate move into this space is a calculated effort to capture market share that top-tier vendors have effectively vacated under compliance pressure.

"Best Open-Source in the US and Europe": A Game of Words

Mistral officially claims Large 4 is "the best open-source model across comprehensive benchmarks in the US and Europe," and that it even surpasses some closed-source frontier models on visual tasks. That sounds impressive — but the framing deserves scrutiny.

In reality, there are very few Western models still committed to a genuine open-source approach. Meta, previously the most formidable open-source competitor, has largely shifted toward closed-source releases. Against that backdrop, claiming the "best open-source in the West" title is more like winning by default. Zoom out to a global view — and bring in Chinese open-source models like GLM, Kimi, or DeepSeek — and the gap becomes apparent.

User testing paints a more nuanced picture than the official benchmarks suggest. Large 4 shows solid planning abilities and code quality, and follows instructions faithfully — perhaps too faithfully. It tends to produce excessively long outputs and can get stuck in loops, earning the nickname "token liberation mode" from some users. In practice, this trillion-parameter model still struggles to keep pace with Chinese models like Qwen3.

Real Progress, But Compute Is the Hard Constraint

Set aside comparisons with Chinese models and look at Large 4 against its own predecessor: the improvement is genuinely striking. Mistral Medium 3.5 scored only 14 on Artificial Analysis, while Large 4 reaches 38 — a massive leap within a single generation.

But Mistral faces a challenge similar to what Chinese companies encounter: insufficient compute. Large 4 was trained on only around 3,800 NVIDIA Blackwell GPUs — a remarkably limited scale.

In a moment rich with symbolism, the same day Large 4 launched, the CEO of German search engine Ecosia announced they were ending their partnership with Mistral, citing that Mistral's models were "already a year behind" and that servers were frequently overloaded. Ecosia holds less than 0.5% of the European search market — and if even this small a company is enough to trigger consistent overload, the scale of Mistral's compute deficit becomes hard to ignore.

NVIDIA Blackwell GPUs (such as the B100 and B200 series) are NVIDIA's latest generation of data center AI accelerator chips as of late 2024, offering significant improvements in training and inference throughput over the previous Hopper architecture (H100/H200). For training a trillion-parameter model, 3,800 Blackwell GPUs is an extremely limited cluster — by comparison, Meta used over 16,000 H100s to train the Llama 3 series, while OpenAI, Google, and other leading labs typically operate training clusters with tens of thousands of GPUs or more. Compute scale directly affects how much training data a model can process, how many training iterations it can run, and ultimately its performance ceiling — as well as its capacity to serve inference at scale. As a European startup, Mistral faces multiple overlapping constraints on GPU access: geopolitical factors, capital limitations, and supply chain challenges. These structural disadvantages are the root cause of the widening gap between Mistral and the leading AI labs in the US and China.

Idealism Can't Substitute for Compute

Mistral's greatest historical advantage has been its commitment to open source, its European identity, and its role as a homegrown alternative to American tech giants. It once gave many people reason to believe that Europe could develop world-class AI without depending entirely on US hyperscalers.

But idealism doesn't translate into compute, and European AI credibility doesn't automatically convert into user adoption. The trillion-parameter scale of Large 4 can't paper over the real gaps in compute infrastructure and real-world engineering. Macron's vision of a "European third way" in AI remains a long road ahead. Large 4 represents genuine, meaningful progress — but for Mistral to truly "reclaim its glory," there is still a very long way to go.