Mistral Large 4 “Le Chonk” Takes the Open AI Crown Outside China

Mistral has officially released the downloadable weights for Mistral Large 4 (ML4), nicknamed “Le Chonk,” positioning it as the highest-scoring open-weight AI model developed outside China, according to independent testing by Artificial Analysis.

Released on October 29, 2026, the trillion-parameter model delivers notable improvements in efficiency, coding performance, and hallucination reduction compared with its October preview.

The biggest breakthrough isn’t raw intelligence. It’s economics.

Mistral has reduced its cost per benchmark task by 63%, from $5.54 to $2.06, while lowering average output-token consumption by 39%. Its coding-agent configuration now ranks second among available open-weight competitors, consuming approximately 47% fewer tokens than the leading GLM-5.3 configuration.

However, the model still trails several Chinese-developed competitors in overall intelligence, factual accuracy, and some coding benchmarks.

For enterprises, Mistral’s strongest selling point may be something benchmarks cannot fully capture: the ability to download, customize, operate, and control a powerful AI model without depending entirely on a third-party provider.

Key Points

1. Mistral claims the top position among open-weight models developed outside China.

Artificial Analysis awarded ML4 a score of 26 on Intelligence Index v5, compared with 25 for its preview and Muse Spark Open.

That puts ML4 narrowly ahead of other open-weight models developed outside China. Globally, however, seven open-weight models perform better, with DeepSeek V4.1 Flash 1017 and GLM-5.3 both scoring 32.

The improvement is incremental rather than a dramatic leap in reasoning capabilities.

2. AI task costs have dropped by an impressive 63%.

The cost per Intelligence Index task declined from $5.54 to $2.06.

Mistral also reduced its API pricing to:

• Input: $0.70 per million tokens.
• Cached input: $0.07 per million tokens.
• Output: $2.10 per million tokens.

Despite these reductions, competitors remain cheaper on the same benchmark. DeepSeek costs $0.60 per task, while MiMo-V2.6-Pro costs just $0.26.

3. Coding performance is becoming a major competitive advantage.

ML4 ranks second on Artificial Analysis’s Coding Agent Index with 48 points, just behind GLM-5.3’s 49.

More importantly, ML4’s agent configuration requires 8.5 million total tokens per task versus GLM-5.3’s 15.9 million.

This represents approximately 47% lower token consumption while delivering a nearly identical benchmark score.

However, these results compare complete coding-agent configurations, not just the underlying models.

4. Hallucinations have declined significantly, but factual accuracy remains a concern.

ML4’s hallucination rate dropped from 41% to 28% on AA-Omniscience v1.1.

That’s a 13-percentage-point improvement.

However, its factual accuracy score remains only 26%, compared with 40% for DeepSeek and 42% for Muse Spark Open.

This makes the distinction between fewer fabricated answers and genuinely accurate answers especially important.

5. Enterprises gain greater control over AI deployment.

Mistral released ML4 under its custom Mistral Ecosystem Open-Weight License.

Organizations can download, modify, further train, and deploy the model internally.

The license has no revenue threshold, but businesses operating model-as-a-service offerings may require separate commercial agreements, depending on their activities and the license’s exceptions.

This gives enterprises an alternative to relying exclusively on hosted AI platforms.

6. Cybersecurity capabilities are promising but not universally superior.

ML4 ranks third among current open-weight models on the Cyber Index, scoring 51.

It achieves 79% on CyberGym-E2E-AA, outperforming DeepSeek’s 76% and matching MiMo.

Mistral also reports a 96.6% refusal rate for malicious cybersecurity prompts.

Nevertheless, ML4 trails competitors on other security evaluations, and the company acknowledges that its capabilities could potentially be misused.

7. Enterprises can customize ML4 for specialized industry applications.

Mistral is targeting industries such as manufacturing, financial services, and cybersecurity.

Organizations can customize the model using Mistral Forge or work directly with Mistral’s engineering and applied science teams.

The company’s broader enterprise focus includes applications involving proprietary documentation, specialized workflows, security controls, and operational continuity.

However, Mistral has not yet published quantified business results specifically attributable to production deployments of ML4.

Key Quotes

On enterprise ownership and independence:

“I don’t want to get disconnected. I want to own my model. Second, I want to customize it to ensure that I derive value. And then third, it’s business continuity.”

— Marjorie Janiewicz, Chief Revenue Officer, Mistral

This captures Mistral’s central enterprise argument: businesses increasingly want control over the AI systems supporting their operations.

On making advanced AI accessible to developers:

“But it’s really for developers at home, all the way to developers in the enterprise.”

— Marjorie Janiewicz

Mistral is positioning ML4 as a model that can support individual developers as well as organizations building specialized enterprise systems.

On cybersecurity and customizable safeguards:

“We believe that the best way for the customers to protect themselves is to have access to those models and be able to establish and customize some of those guardrails themselves.”

— Marjorie Janiewicz

Mistral argues that enterprises should be empowered to establish security controls appropriate to their own environments rather than relying exclusively on provider-defined restrictions.

On industry specialization:

“The models we build are purpose built for the industries we serve.”

— Marjorie Janiewicz

The company emphasizes adaptation to specialized business environments rather than competing solely on general-purpose benchmarks.

On decreasing enterprise AI expenses:

“A lot of our customers, particularly over the last six months, are really beginning to see their bills decreasing.”

— Marjorie Janiewicz

While customers are reportedly seeing lower bills, Mistral has not yet provided quantified operational savings attributable specifically to ML4.

Implications

1. AI competition is increasingly about efficiency, not just intelligence.

Mistral’s biggest achievement is reducing the cost of completing benchmark tasks while delivering competitive performance.

The 63% cost reduction and 39% decrease in average output-token usage demonstrate that improvements in efficiency can be commercially meaningful even when intelligence scores rise only slightly.

For enterprise buyers, measuring the cost of successfully completed work may become more useful than comparing model rankings alone.

2. Open-weight models offer enterprises an alternative to provider dependence.

With downloadable weights and customization rights, ML4 gives organizations more options for maintaining control over their AI infrastructure.

This could be particularly relevant to businesses prioritizing proprietary information, continuity of operations, and specialized deployment requirements.

However, operating a model independently also introduces infrastructure, maintenance, and oversight costs that API benchmarks do not capture.

3. Coding agents could become more economically practical.

ML4’s competitive coding-agent score, combined with substantially lower token usage than the leading configuration, strengthens the argument for using AI agents in software development.

The important qualification is that the results depend on the complete agent setup. Enterprises still need to establish whether those efficiencies translate into reliable performance on their own development tasks.

4. Lower hallucination rates do not eliminate the need for verification.

Although ML4 generates fewer fabricated responses on the tested benchmark, its factual accuracy remains substantially below several competitors.

Businesses considering the model for knowledge-intensive workflows must distinguish between an AI system that avoids unsupported claims and one that consistently supplies correct information.

Human oversight and task-specific evaluation remain important.

5. Customization creates opportunities, and additional responsibilities.

Mistral’s approach allows companies to shape models around proprietary data and internal processes.

That flexibility could make ML4 attractive for industries with demanding operational and security requirements.

But greater control also shifts more responsibility to the deploying organization, particularly when it comes to access restrictions, guardrails, testing, and cybersecurity.

6. The biggest unanswered question is measurable business value.

Mistral has demonstrated meaningful improvements across public benchmarks, including coding, efficiency, hallucination reduction, and selected enterprise evaluations.

What remains unproven is how those improvements translate into customer-specific productivity gains, revenue improvements, or return on investment.

The model’s ultimate enterprise value will depend on whether organizations can turn its lower costs, customization capabilities, and competitive performance into reliable outcomes in real operating environments.

The bottom line: Mistral Large 4 doesn’t win the global intelligence race. Instead, it strengthens a different proposition: powerful, customizable AI that businesses can operate with greater independence and significantly improved efficiency. Its success will depend on whether that combination proves more valuable to enterprise customers than cheaper or higher-scoring alternatives.

Source: https://venturebeat.com/technology/mistral-large-4-le-chonk-arrives-as-the-top-open-model-outside-china-with-improved-coding-performance-and-63-lower-cost-per-task