Mistral Large 4

mistral.ai

October 21, 2026

11 min read

🔥🔥🔥🔥🔥

62/100

Summary

Mistral has launched a public-preview API for Mistral Large 4, a natively multimodal mixture-of-experts model with 1 trillion total parameters and 52 billion active parameters. The company plans to release its open weights by the end of the month. Mistral Large 4 was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral-operated European data centers, using multilingual data spanning more than 160 languages. The preview API is available through Mistral Studio, priced at $1.36 per million input tokens and $4.18 per million output tokens. Mistral says the model is its largest and most capable release, targeting coding, tool-using agents, cybersecurity, finance, law, science, and visual understanding. It reported scores of 82% on a vulnerability reproduction-and-patching test, 93% on Cybench security challenges, 61.7% on DeepSWE v1.1, 59.9% on AutomationBench, and 42% on Dense 200 visual grounding. Mistral also said ML4 resists 93.3% of attacks in Lakera’s B3 AI Security Benchmark. Before the weight release, Mistral is red-teaming the model with cybersecurity leaders, vetted partners, and state authorities that will receive reduced moderation and expanded cyber capabilities. The company says ML4 can be self-deployed in private clouds or on premises and will underpin specialized future Mistral models. Its development is funded in part by Mistral’s €3 billion Series D round.

Key Takeaways

Mistral Large 4 is a 1 trillion-parameter multimodal mixture-of-experts model with 52 billion active parameters, and Mistral plans to publish its weights by the end of the month.
Mistral trained ML4 on 3,800 Nvidia Grace Blackwell GPUs in its European data centers and offers its preview API through Mistral Studio.
Mistral reported that ML4 scored 82% on a vulnerability reproduction-and-patching test and solved 93% of Cybench’s 40 cybersecurity challenges.
Mistral reported scores of 61.7% on DeepSWE v1.1 for software engineering, 59.9% on AutomationBench for business workflows, and 42% on Dense 200 visual grounding.
Mistral is red-teaming ML4 with cybersecurity leaders, vetted partners, and state authorities before releasing the weights, including access with reduced moderation and expanded cyber capabilities.

What the discussion said

The thread treated Mistral Large 4 less as a clean launch than as a pricing-and-positioning problem. Several commenters see a genuinely competitive multimodal model: its visual work is said to outpace a DeepSeek Flash variant, it feels fast, and its large sparse architecture is noteworthy because it reportedly reaches near-frontier territory with far less training infrastructure. The planned open-weight release also keeps interest alive, especially for people hoping to test or self-host it rather than accept headline benchmarks. But the dominant practical question was why anyone should pay for it when DeepSeek appears cheaper and stronger on broad comparisons, especially coding. Some readers identified concrete reasons: European hosting and GDPR coverage, clearer training opt-out, data-retention constraints, and inclusion in an existing Mistral subscription. Others argued model choice cannot be reduced to scorecards because behavioral fit varies by task. That defense met skepticism from an adversarial programming evaluation where the model reportedly performed only around a lower-tier competitor. Reliability also drew attention after reports of unexplained Chinese characters appearing in German responses. Readers connected the symptom to possible quantization damage or cross-model artifacts, while one speculated about distillation without evidence. Overall, commenters grant that Mistral remains credible, but want real-world tests and a sharper value proposition before calling it a winner.

Where opinion split

The sharp dispute is whether Mistral Large 4 has a compelling reason to exist beside DeepSeek. Supporters argue that multimodal strengths, jurisdictional and privacy requirements, and task-specific behavior can outweigh aggregate benchmark comparisons; critics say those benefits do not justify roughly tenfold higher cost when it trails on much of the visible scoreboard, particularly coding.

Read original article

View story →

Coverage & discussions

Hacker News · mistral.ai

Mistral Large 4 ↗

Community discussions

Hacker News discussion ↗

Reported comments: 92 · 12 saved in sample

Analysis ready

Community Sentiment

Mixed

Positives

Its visual and multimodal performance reportedly beats a DeepSeek Flash competitor, giving Mistral a plausible niche beyond raw coding scores.
Near-frontier results from substantially less training infrastructure impressed readers, suggesting strong capability need not remain exclusive to the largest AI labs.
European governance, GDPR coverage, and a straightforward training opt-out make it a practical model choice where data handling rules rule out foreign providers.
Fast responses and a forthcoming open-weight release give developers reasons to test it directly rather than judge solely by launch benchmarks.
Several commenters stressed that model behavior is use-case dependent, so a weaker aggregate ranking can still conceal a better fit for a specific workflow.

Concerns

At around ten times DeepSeek's price while looking weaker on broad comparisons, the model's value proposition struck many readers as hard to defend.
An adversarial programming-language benchmark reportedly placed it near Kimi K3, undercutting claims that it is a serious coding-model upgrade.
Reports of Chinese characters leaking into German chats raise reliability concerns, whether caused by model behavior, aggressive quantization, or another deployment artifact.
The model's visual edge does not offset its coding deficit for readers whose main AI workload is software development.