Mistral Large 4 (le Chonk): 1-Trillion-Parameter Open-Weight Frontier Model with MoE Architecture, Cybersecurity Leadership, and Agentic Coding

Executive Summary

On October 6, 2026, Mistral AI launched a public preview of Mistral Large 4 — internally nicknamed "le Chonk" — representing the largest open-weight model release to date. At 1.05 trillion total parameters with 52 billion active parameters (via granular Mixture-of-Experts architecture), it is Mistral's most capable model by a wide margin and the first open-weight model to approach frontier closed-model performance on critical enterprise workloads.

The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters — a significant milestone for European AI sovereignty. It is natively multimodal (text + image), supports a 1M-token context window, and demonstrates exceptional performance across cybersecurity, agentic coding, finance, law, and visual grounding. On cybersecurity benchmarks, it leads all open-weight models by a wide margin and outperforms several closed models that refuse to perform the same tasks due to safety filters.

At introductory pricing of $0.68 per million input tokens and $2.09 per million output tokens (50% off the standard $1.36/$4.18 rates), with full weights promised by October 31, 2026, Mistral Large 4 represents a strategic shift in the open-weight frontier. It is the first milestone funded by Mistral's €3 billion Series D — the largest equity round ever raised by a European technology company — and signals a sustained push toward sovereign, auditable AI infrastructure.

---

1. What Is Mistral Large 4?

1.1 The Scale

Mistral Large 4 is a natively multimodal model with a granular Mixture-of-Experts (MoE) architecture:

Property	Value
Total parameters	1.05 trillion
Active parameters per token	52 billion
Vision encoder	1.6 billion parameters
Context window	1 million tokens
Input modalities	Text, Image
Output modalities	Text
Training infrastructure	3,800 NVIDIA Grace Blackwell GPUs
Training location	European datacenters (Mistral-owned)
Languages in training data	160+ (including all EU official languages)
License	Open-weight (details TBD with weight release)
Weight release ETA	October 31, 2026

The MoE architecture means that while the model contains 1.05T parameters total, only 52B are activated per token — a ~5% activation rate that enables frontier-scale capability with manageable inference costs. The 1.6B-parameter vision encoder is integrated natively, not bolted on, enabling true multimodal reasoning rather than sequential text-then-vision processing.

1.2 European AI Sovereignty

A defining characteristic of Mistral Large 4 is its European sovereignty story:

"ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own datacenters in Europe. The public preview is served on that same infrastructure."

This matters for several reasons:

• Data residency: Training and inference occur entirely within European jurisdiction
• Legal compliance: Operates under European law, independently of US digital service providers
• Supply chain control: Mistral owns the infrastructure, reducing dependency on cloud providers
• Multilingual foundation: Training data spans 160+ languages, including every official EU language

The model will be available across multiple regions, including a European deployment that Mistral operates end-to-end — a feature particularly relevant for government, defense, and regulated industry customers.

1.3 The Open-Weight Promise

Mistral Large 4 is currently in public preview via the Mistral Studio API. Full weights will be released by October 31, 2026, along with:

• Detailed model architecture documentation
• Additional benchmarks
• Post-training methodology details

The planned Hugging Face artifact identifier is Mistral-Large-4.0-1T05-A52B. At approximately 1TB of weights in FP8 format, self-deployment will require significant infrastructure — roughly one node of B200s or H200s based on Mistral Large 3's requirements scaled to the larger parameter count.

---

2. Architecture: Granular Mixture-of-Experts

2.1 MoE Design

Mistral Large 4 uses a granular Mixture-of-Experts architecture — a refinement over the coarse MoE designs seen in earlier models. The key characteristics:

• 1.05T total parameters distributed across many expert subnetworks
• 52B active parameters per token (approximately 5% of total)
• Dynamic routing that selects which experts process each token based on content
• 1.6B parameter vision encoder integrated as a native modality

The granular approach allows the model to maintain massive total capacity while keeping per-token compute manageable. This is the same architectural family as Mixtral 8x22B and DeepSeek V3/V4, but scaled to an order of magnitude larger.

2.2 Training Methodology

Mistral describes the training as "from scratch" — not a fine-tune of an existing model. The training pipeline includes:

• Supervised fine-tuning (SFT) — initial alignment on curated datasets
• Reinforcement Learning (RL) — extensive post-training using Mistral's custom RL library
• Enterprise collaboration — trained with input from leading enterprises in finance, engineering, manufacturing, logistics, pharmaceuticals, science, shipping, and public sector

The RL infrastructure is notable:

"At our current scale (3k GPUs), a single training run produces roughly 33 billion tokens per day, of which around 16 billion are trainable completion tokens after filtering and masking."

The RL library supports composable environments — a single training run can combine single-turn chat, complex scientific problem solving, safety alignment, factuality, and long-horizon tool use. Verification uses reward models, unit tests, LLM judges, and static checks.

---

3. Benchmark Performance

3.1 Cybersecurity: Leading Open-Weight Models

Mistral Large 4's most distinctive achievement is in cybersecurity, where it leads open-weight models by a wide margin and outperforms several closed models on specific tasks:

Benchmark	Mistral Large 4	Best Closed Model	Notes
Cybench	93%	—	40 security competition exercises
Vulnerability reproduction + patch	82%	Near zero (Claude Opus 5.5, GPT-6 Astra)	Closed models refuse the task
AA Cyber Index	Top 5 globally	—	Independent evaluation

The vulnerability reproduction result is particularly significant. Several leading closed models (Claude Opus 5.5, GPT-6 Astra) score near zero on this test because their safety filters refuse to reproduce real vulnerabilities — even though understanding and reproducing vulnerabilities is essential for defense:

"Defending software often starts with proving that a flaw is real, exactly the kind of work safety filters in closed models can block."

This creates a paradox: closed models with strong safety filters are less useful for cybersecurity defense, while threat actors can jailbreak those same models for offensive purposes. Mistral Large 4, as an open-weight model, allows organizations to run advanced security work under their own policies.

3.2 Agentic Coding

On coding benchmarks, Mistral Large 4 demonstrates strong performance across software engineering and terminal workflows:

Benchmark	Mistral Large 4	DeepSeek V4 Pro 0813	Qwen3.8 Max	GPT-6.1 Sol	Claude Opus 5.5
DeepSWE v1.1	61.7%	—	—	75.2%	—
SWE-Atlas-QnA	59.4%	—	—	—	—
Terminal-Bench 4.0	28.3%	—	—	—	—
Coding Agent Index	49.8%	Below	Below	—	—
Vibe Code Bench v1.1	78.4%	—	—	—	—

The Coding Agent Index score of 49.8% places it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. On the Vibe Code Bench v1.1 (evaluated by vals.ai), it achieves 78.4%.

3.3 Agentic Workflows

On general-purpose agentic workflows, Mistral Large 4 performs strongly:

Benchmark	Mistral Large 4	Kimi K3	MiMo-V2.6-Pro	DeepSeek V4 Pro
AutomationBench	59.9%	Below	Below	Below
AA-Briefcase	1,393 Elo	—	—	Below
Finance Agent v2	54.7%	—	—	—

On AutomationBench — 657 business workflows across Gmail, Google Sheets, Slack, and Salesforce — it scores 59.9%, ahead of Kimi K3, MiMo-V2.6-Pro, and DeepSeek V4 Pro. On AA-Briefcase, which evaluates long-horizon knowledge work (spreadsheets, slides, PDFs), it reaches 1,393 Elo.

3.4 Knowledge Work: Finance and Law

Mistral Large 4 was evaluated by third-party evaluators (vals.ai) on legal and financial tasks, finding the model exceeds GPT-6 Astra in both cases:

Domain	Benchmark	Mistral Large 4	GPT-6 Astra	Notes
Legal	HarveyAI Legal Agent	Leads all open-source models	Exceeded by ML4	Third-party eval (vals.ai)
Finance	FinWorkBench	Exceeds GPT-6 Astra	Below ML4	Real-world spreadsheet tasks

This is a notable result — outperforming GPT-6 Astra on specialized enterprise workloads while being an open-weight model at a fraction of the cost.

3.5 Multimodal & Visual Grounding

Mistral Large 4 demonstrates strong multimodal capabilities, particularly in visual grounding:

Benchmark	Mistral Large 4	GPT-6 Astra	Notes
Dense 200 (visual grounding)	42%	41%	Surpasses closed model
AA-MMMU-Pro	76.4%	—	Multimodal understanding

The model can combine visual grounding with agentic capabilities: inspecting gigapixel satellite imagery for disaster response, analyzing engineering drawings, retrieving evidence from PDFs, and scanning massive geospatial images.

3.6 Science and Math

Domain	Achievement
SciCode-Verified	State-of-the-art among open-weight models
Hartree–Fock simulation	Generates full simulation in one shot (multi-step chemistry)
Formal math	Stronger than GLM-5.3 in human evaluations
Applied math	Sustains long, domain-specific tasks including frontier theoretical physics

3.7 Reasoning & Knowledge

Benchmark	Score	Source
AA-LCR (Long Context Reasoning)	81.3%	Artificial Analysis
AA-HLE (Humanity's Last Exam)	35.0%	Artificial Analysis
AA Intelligence Index	38.4%	Artificial Analysis
AA-Omniscience Accuracy	25.8%	Artificial Analysis
GDPval-AA	1,424	Artificial Analysis

3.8 Human Evaluation

A blind human evaluation by Surge AI rated model outputs on a 1–5 scale:

Model	Score
Claude Opus 5	4.22
Mistral Large 4	3.74
GLM-5.3	3.60
Kimi K3	3.59
GLM-5.2	3.40

Mistral Large 4 ranked second of five models, ahead of all major Chinese open-weight models and behind only Claude Opus 5.

---

4. Model Safety

4.1 Prompt Injection Resistance

Mistral Large 4 has saturated Mistral's benchmarks on robustness to indirect prompt injections:

Benchmark	Mistral Large 4	Notes
B3 AI Security Benchmark (Lakera)	93.3%	No higher scores among competitors
KORA Benchmark	1.691 / 2.0	Highest measured score among OSS models

4.2 Cybersecurity Refusal Calibration

Despite strong performance on cyber benchmarks, the model maintains high refusal rates on malicious requests:

"The average refusal rate of the model on cyber prompts from JailbreakBench, StrongREJECT, and AgentHarm is higher than all OSS models."

This represents a careful balance: strong capability for legitimate security research while maintaining robust refusals for malicious requests.

4.3 Red-Team Access

Until the full weight release, Mistral is red-teaming the model in real-world settings with:

• Cybersecurity leaders
• Vetted partners
• State authorities

These partners access the same model with reduced moderation and expanded cyber capabilities — a controlled environment for testing the model's security boundaries.

---

5. Pricing & Cost Economics

5.1 Introductory Pricing

Mistral Large 4 is available at 50% introductory pricing through the end of 2026:

Price per 1M tokens	Standard	Introductory (50% off)
Input	$1.36	$0.68
Cached input	$0.14	$0.07
Output	$4.18	$2.09

The cached input rate of $0.07/M is among the lowest of any frontier model — a 95% discount from standard input pricing.

5.2 Cost Comparison

Model	Input ($/1M)	Output ($/1M)	Cached Input	Context	Open Weights
Mistral Large 4	$0.68	$2.09	$0.07	1M	Yes (Oct 31)
GPT-6.1 Sol	$2.00	$10.00	$0.10	1.05M	No
Grok 4.7	$2.00	$6.00	$0.50	500K	No
Sonnet 5.5	$2.00	$10.00	$0.10	1M	No
Haiku 5.5	$0.10	$0.50	$0.01	1M	No
GPT-6 Luna	$0.10	$0.50	$0.01	1.05M	No

At introductory pricing, Mistral Large 4 is ~3× cheaper on input and ~5× cheaper on output than GPT-6.1 Sol, while offering open weights for self-deployment.

5.3 Batch Pricing

Batch processing reduces prices by an additional 50%, making high-volume workloads even more economical.

---

6. Deployment & Access

6.1 Availability Channels

Channel	Status	Notes
Mistral Studio API	✅ Available	Public preview, mistral-large-4-0
Mistral Studio Dashboard	✅ Available	Usage tracking, metrics
Vibe (web/CLI/IDE/mobile)	✅ Available	Personal AI agent interface
Hugging Face	📅 Oct 31, 2026	Full weight release
Self-deployment	📅 Oct 31, 2026	Private cloud or on-premise
European deployment	✅ Available	End-to-end Mistral-operated
Mistral Forge	✅ Available	Customization and RL environment

6.2 Self-Deployment Considerations

The 1.05T parameter count means self-deployment requires significant infrastructure:

• ~1TB of weights at FP8 precision
• Minimum hardware: One node of B200s or H200s (based on Large 3 scaling)
• Inference optimization: MoE architecture enables efficient inference with 52B active parameters
• Use cases: Sovereign AI for government, defense, regulated industries

6.3 Code Example: API Usage

import os
from mistralai import Mistral

client = Mistral(api_key=os.getenv("MISTRAL_API_KEY"))

response = client.chat.complete(
    model="mistral-large-4-0",
    messages=[
        {
            "role": "user",
            "content": "Analyze this engineering drawing and identify any "
                       "potential structural issues in the load-bearing components."
        }
    ],
    # Include image for multimodal analysis
    # images=[{"url": "https://example.com/drawing.png"}],
)

print(response.choices[0].message.content)

---

7. Connection to Prior Research

7.1 The October 2026 Frontier Landscape

Mistral Large 4 arrives in the most active period of the frontier AI race:

• Gpt 6 1 Sol Near Astra Performance Fifth Cost Agentic Coding Computer Use Professional Work 2026 10 09 — OpenAI's near-Astra mid-tier model
• Claude Haiku 5 5 Cheapest Fastest Small Model Adjustable Effort 1m Context Subagent Browser Use 2026 10 08 — Anthropic's $0.10/M small model
• Grok 4 7 Frontier Coding Knowledge Work New Safeguard Stack Longer Rl 2026 10 06 — xAI's frontier coding model with new safeguards
• Qwen3 8 Omni Flash Native Omnimodal Agentic Media 1m Context 2026 10 05 — Qwen's native omnimodal model
• Gemini 4 Argon Frontier Intelligence Cyber Defense 1m Output Tokens Limited Release 2026 09 30 — Google's cyber-defense model
• This article — Mistral Large 4 with 1T parameters and open-weight sovereignty

7.2 The Open-Weight Frontier Race

Mistral Large 4 represents the largest open-weight model released to date, significantly larger than previous open-weight flagships:

Model	Total Parameters	Active Parameters	Architecture	Open Weights
Mistral Large 4	1.05T	52B	Granular MoE	Oct 31, 2026
DeepSeek V4	~763B	~39B	MoE	Yes
Qwen3.8 Max	~235B	~235B	Dense	Yes
Mixtral 8x22B	176B	39B	MoE	Yes
Mistral Large 3	~128B	~128B	Dense	Yes

7.3 The Cybersecurity Paradox

Mistral Large 4 highlights a paradox in the frontier landscape that connects to Grok 4 7 Frontier Coding Knowledge Work New Safeguard Stack Longer Rl 2026 10 06 and Gemini 4 Argon Frontier Intelligence Cyber Defense 1m Output Tokens Limited Release 2026 09 30:

• Closed models (Claude Opus 5.5, GPT-6 Astra) refuse vulnerability reproduction tasks due to safety filters
• Open-weight models (Mistral Large 4) allow organizations to run security work under their own policies
• Threat actors can jailbreak closed models for offensive purposes regardless

This creates a security asymmetry where defenders are more constrained than attackers — a problem that open-weight models with self-deployment can address.

7.4 The Cost-Performance Continuum

Mistral Large 4 fills a gap in the cost-performance spectrum:

Tier	Model	Input ($/1M)	Output ($/1M)	Positioning
Flagship closed	GPT-6 Astra	$10.00	$50.00	Maximum capability
Mid-tier closed	GPT-6.1 Sol	$2.00	$10.00	Near-flagship
Open-weight frontier	Mistral Large 4	$0.68	$2.09	Frontier open
Small-model closed	Haiku 5.5	$0.10	$0.50	High-volume
Small-model closed	GPT-6 Luna	$0.10	$0.50	High-volume

Mistral Large 4 offers frontier open-weight capability at mid-tier closed-model pricing, with the additional benefit of self-deployment.

---

8. Key Takeaways

1. 1.05T parameters is a new scale for open-weight models. Mistral Large 4 is the largest open-weight model released to date, with a granular MoE architecture that activates only 52B parameters per token.

2. Cybersecurity leadership is the standout achievement. Scoring 93% on Cybench and 82% on vulnerability reproduction — where closed models like Claude Opus 5.5 and GPT-6 Astra score near zero due to safety refusals — makes ML4 uniquely valuable for defense organizations.

3. European AI sovereignty is a strategic differentiator. Trained and served entirely in European datacenters under European law, with support for 160+ languages including all EU official languages.

4. Open weights by October 31. The promise of full weight release — along with detailed architecture documentation — enables self-deployment for organizations needing sovereign, auditable AI.

5. Enterprise workloads are a sweet spot. Outperforming GPT-6 Astra on finance and legal benchmarks (per vals.ai third-party evaluation) while being significantly cheaper.

6. The €3B Series D fuels sustained investment. This is the first milestone of what Mistral describes as a long-term push, with more compute coming online and the RL training run still in flight.

7. Visual grounding surpasses closed models. On Dense 200, ML4 scores 42% vs. GPT-6 Astra's 41%, demonstrating that open-weight models can lead even on closed-model strengths.

8. The safety balance is carefully calibrated. 93.3% resistance to prompt injection attacks (B3 benchmark) while maintaining high refusal rates on malicious cyber requests — all while enabling legitimate security research.

---

9. Future Directions

9.1 What to Watch

1. Weight release (October 31): The full weight release will determine self-deployment feasibility and enable the community to fine-tune, quantize, and extend the model.

2. Architecture details: Mistral has promised more details on the granular MoE architecture, which could influence the design of future open-weight models.

3. RL training continuation: The reinforcement learning run behind the preview is "still in flight" with "substantial headroom ahead" — expect the model to improve significantly between now and the weight release.

4. Specialized models: Mistral states that ML4 "will serve as the foundation for a new generation of specialized and optimized Mistral models" — expect domain-specific variants.

5. Competitor response: The 1.05T parameter scale sets a new bar for open-weight models. Expect DeepSeek, Qwen, and others to respond with larger or more efficient architectures.

6. Enterprise adoption: The combination of European sovereignty, cybersecurity leadership, and open weights could drive adoption in government, defense, and regulated industries.

7. Infrastructure scaling: The €3B Series D is already being used to scale compute capacity in European datacenters — the pace of infrastructure growth will determine how quickly Mistral can iterate.

8. License terms: The specific license for the open weights will determine what organizations can and cannot do with the model — particularly around derivatives and commercial use.

---

10. References & Resources

Official Sources

• Introducing Mistral Large 4 — Official announcement with benchmarks, architecture overview, and capabilities deep-dive
• Mistral Large 4 Model Page — Model specifications, pricing, and features
• Mistral Studio — Public preview API access
• Mistral Forge — Customization and RL environment
• Hugging Face (planned) — Weight release (ETA October 31, 2026)

Independent Benchmarks

• BenchLM: Mistral Large 4 — Aggregated benchmark scores
• Artificial Analysis: Mistral Large 4 — Independent evaluation
• Vals AI: Finance Agent v2 — Financial task evaluation
• Vals AI: Vibe Code Bench v1.1 — Coding evaluation
• B3 AI Security Benchmark — Prompt injection resistance
• KORA Benchmark — Responsible engagement evaluation

Related Research in This Journal

• Gpt 6 1 Sol Near Astra Performance Fifth Cost Agentic Coding Computer Use Professional Work 2026 10 09
• Claude Haiku 5 5 Cheapest Fastest Small Model Adjustable Effort 1m Context Subagent Browser Use 2026 10 08
• Grok 4 7 Frontier Coding Knowledge Work New Safeguard Stack Longer Rl 2026 10 06
• Qwen3 8 Omni Flash Native Omnimodal Agentic Media 1m Context 2026 10 05
• Gemini 4 Argon Frontier Intelligence Cyber Defense 1m Output Tokens Limited Release 2026 09 30

---

This article was researched using official Mistral AI sources (Mistral Large 4 announcement page, model documentation, and pricing page) supplemented by benchmark data from BenchLM, Artificial Analysis, and vals.ai. All benchmark figures from Mistral's launch post are cited directly with appropriate caveats about vendor-reported results. Independent scores from Artificial Analysis are noted where available.