Mistral Releases Large 4 Preview: Trillion Parameters Aimed at Complex Enterprise Workflows

Mistral has launched the public preview API for its flagship model, Mistral Large 4 (codenamed ML4, nicknamed "le Chonk" internally). The company announced plans to open model weights by the end of the month (weights have not yet been released, and specific licensing details have not been provided), while introducing its European self-built data center powered by an autonomous infrastructure of 3,800 Nvidia Grace Blackwell GPUs.

The model features 1.05 trillion total parameters (1.05T) with a sparse Mixture-of-Experts (MoE) architecture, activating roughly 49 billion parameters (49B) per token forward pass; in terms of product positioning, it consolidates instruction following, reasoning, and multimodal agent capabilities into a single model. Mistral highlights enterprise controllability as a key focus, targeting scenarios with stringent data and operational policy requirements such as real-world cybersecurity, complex software engineering, long-horizon agentic tool use, and specialized financial and legal workflows.

Architecture and Scale: Why "le Chonk"?

In its model card and release announcements, Mistral playfully refers to ML4 as "le Chonk" (an internet slang term describing a chunky, hefty animal). Looking at its specifications, it is indeed a behemoth, with very explicit architectural trade-offs:

• 1.05 trillion total parameters, 49 billion active parameters: It uses a granular Mixture-of-Experts architecture (Granular MoE). The distinction between total and active parameters is that total parameters constitute the model's complete knowledge network, meaning a deployment system must still store the full set of weights; active parameters represent the specific expert modules selected by gating routers to participate in floating-point operations during each token generation step. While MoE reduces the number of active parameters engaged in computing each token, the server must still hold the massive 1.05T parameter weights—reducing active computation does not shrink the overall memory footprint; actual latency and serving costs depend on hardware, precision, and serving configurations.
• Native vision encoder and 1M context: The model integrates a native vision encoder of roughly 1.6 billion parameters (1.6B) and supports a 1-million-token context window (1M Context). These are two distinct specifications: a 1M context accommodates lengthy documents or extensive multi-turn histories, but when handling ultra-high-resolution satellite imagery or dense mechanical blueprints, the model cannot magically resolve every minute detail in a single pass without external tools like viewport zooming and localized cropping to enable cross-level inspection.
• Multilingualism and European sovereign operations: Training data spans more than 160 languages, including all official languages of the European Union. Mistral partnered with enterprises across finance, engineering, logistics, and healthcare to train ML4, utilizing the same training, fine-tuning, and reinforcement learning platform provided to enterprise customers (Mistral Forge). This bridges general foundation model research with domain-specific enterprise customization: the exact training tools offered to clients are used on Mistral's own flagship model. The model was trained from scratch in Europe on 3,800 Nvidia Grace Blackwell GPUs in Mistral's proprietary data center, and the preview version is hosted on this same infrastructure. Mistral emphasizes end-to-end sovereign operations under European jurisdiction without reliance on external cloud providers, though self-hosted on-premises enterprise deployments must await the actual release of weights and licensing terms.

Cybersecurity in Practice: When Legitimate Exploit Reproduction Is Blocked, How Do Open Weights Help?

In cybersecurity operations, defensive teams frequently face a practical dilemma: patching a vulnerability often requires reproducing its trigger conditions in a controlled environment first. However, certain commercial cloud models enforce strict safety guardrails that, when encountering exploit signatures or penetration payloads, may trigger provider-level safety refusals (Safety Refusal), halting legitimate reproduction and incident response tasks. Mistral stresses that losing access to model capabilities in the middle of resolving a security incident can itself constitute a severe operational hazard: investigations, reproductions, and remediation workflows that already depend on the model may grind to an abrupt halt. The value of autonomous operations therefore includes operational continuity during emergencies—enabling organizations to schedule and audit security tasks under their own governance policies rather than depending on external provider decisions. Open weights and self-hosting remain planned, with the current public preview hosted directly by Mistral.

In the third-party evaluation data from Artificial Analysis Cyber Index shared by Mistral, model performance across cybersecurity tasks is highlighted:

• Composite metric (Cyber Index): In this composite benchmark aggregating multiple security evaluations, the original chart plots task success rate alongside safety refusal rate. ML4 achieved a 50% success rate; by comparison, Qwen3.8 2.4T A95B exhibited a 63% safety refusal rate (with a 13% success rate), while Claude Opus 5.5 and GPT-6 Astra recorded safety refusal rates of 38% (with success rates of 29% and 33%, respectively). Thus, the benchmark reflects both problem-solving capability and the model's willingness to execute the task.
• Real-world exploit reproduction and patching (CyberGym-E2E): As a dedicated sub-task within the Cyber Index, this benchmark evaluates reproducing vulnerabilities in real-world open-source software and generating patches. ML4 scored 82% on this sub-task, outperforming MiMo-V2.6-Pro (79%), GLM-5.3 (74%), and Kimi K3 (58%). Mistral notes that Claude Opus 5.5 and GPT-6 Astra scored near zero on this benchmark due to upfront refusals to reproduce exploits; a low score resulting from refusal is fundamentally different from attempting the task and failing.
• 40-challenge CTF competition (Cybench): Across the 40 challenges in Cybench, ML4 solved 93% of the problems, ranking among the stronger reported open-weight models. Furthermore, in an official internal CTF Speedrun demo featuring 19 internal security challenges, ML4 solved 18 (in the same comparison, GLM-5.2 solved 16, Kimi K3 solved 15, GLM-5.3 solved 14, DeepSeek V4.1 Flash solved 14, and DeepSeek V4 Pro solved 10); this 19-challenge internal demo is a distinct task set from the 40-challenge Cybench benchmark.

Reverse Engineering Is Not a "One-Click Wonder": The 56-Turn Trajectory in Practice

In an official demonstration analyzing a suspicious unknown binary file sample.bin, the model progressed through 56 trajectory turns (trajectory turns rather than merely 56 tool calls), demonstrating an iterative, error-correcting investigative process:

1. Static feature analysis: The model first invoked tools to inspect the file format, import tables, and string listings.
2. Decompilation and payload decoding: It then decompiled suspicious routines and read raw bytes, extracting and decoding an 892-byte encrypted payload (Payload).
3. Impeded API hash resolution: The model attempted to resolve Windows API hashes to identify system call behaviors, but encountered repeated lookup failures along the way.
4. Pivot to rule authoring and validation: After hash lookups stalled, the model pivoted to drafting YARA signature rules based on the recovered logic; once compiled, the rule successfully detected the sample and confirmed a 0 false-positive rate across a test corpus of 683 benign files.
5. Evidence grounding and final delivery: Upon submitting the final analysis report, the background automated auditor rejected the report three consecutive times for "claims lacking supporting evidence." The model resolved each missing citation using Claim IDs returned by the auditor, ultimately delivering a 20-claim report comprising 17 verified conclusions, 2 reasonable inferences, and 1 unconfirmed claim.

This ability to iterate, fail, and self-correct across extended tool chains is a critical prerequisite for production security agents. Supporting legitimate exploit reproduction does not imply removing safety boundaries: the model continues to enforce refusals against malicious attacks. In Mistral's reported malicious web and cyber request refusal evaluations (aggregating JailbreakBench, StrongREJECT, and AgentHarm), ML4 achieved an average refusal rate of 95.3%; on the Lakera B3 benchmark assessing defense against indirect prompt injection (hidden instructions embedded in web pages or documents designed to hijack agents), ML4 registered a 93.3% attack resistance rate, tying GLM-5.3 for the top score; in the KORA benchmark measuring responsible interaction, ML4 scored 1.691 (out of 2). Furthermore, the "reduced filtering and expanded cybersecurity capabilities" release mentioned by Mistral is restricted to vetted partners, security organizations, and government red teams for joint exercises, and does not mean the public API or upcoming standard weights are free of safety constraints.

Code and Agent Collaboration: Actual Benchmark Performance

For software engineering and agentic workflows, Mistral shared several benchmark evaluations. On the Coding Agent Index—a composite metric combining bug fixing, terminal operations, and repository Q&A—ML4 scored 49.8%, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. Individual benchmark results are as follows:

Looking across these numbers, ML4's performance varies across code tasks: while it leads several open-weight models on DeepSWE and command-line tasks, it trails GLM-5.3 on Terminal-Bench and sits slightly below DeepSeek V4 Pro and Qwen3.8 Max on repository Q&A. Cross-model comparisons must also account for evaluation harnesses (Harness) and tooling setups: some models utilized different coding tools, and certain Artificial Analysis scores stem from private evaluations conducted before the harness was publicly released, meaning score differences do not directly translate into production efficacy.

From Medium 3.5 to ML4: How Much Have Workflow Capabilities Improved?

Beyond comparisons with competitors, the generation-over-generation evolution within Mistral's own lineup is noteworthy. Official charts place Medium 3.5 and ML4 side by side on the same task suite:

Workflows increased from 6.3% to 59.9%, a gain of 53.6 percentage points. In the same comparison, Finance Agent scores climbed from 32.1 to 54.7, accounting spreadsheets from 36.7 to 67.4; scientific programming pass@1 improved from 70% to 91.8%, while chart Q&A moved from 55.4 to 63.1.

Across this test suite, the new model completes substantially more cross-application tasks; financial spreadsheets and scientific programming also show major gains, while chart Q&A shows a more modest increase. For existing Mistral users, these shifts more directly address the question: "When upgrading to the new model, which workloads are worth re-testing?" These scores do not isolate why the prior model failed, nor do they translate directly into execution speed or attribute gains exclusively to RL.

Domain-Specific Scenarios: Legal, Financial, and Multimodal Engineering Drawings

When general-purpose foundation models enter specialized verticals, the core challenge is whether they can execute rigorous professional processes. In legal liability assessment, financial auditing, and industrial engineering drawing analysis, evaluations typically center on concrete domain tasks.

1. Legal and Financial Workflows: Task-Oriented Execution

In industry evaluations organized by third-party benchmark provider vals.ai, assessments measure the ability to handle professional workflows rather than broad trivia Q&A:

• Legal Agent evaluation (Harvey's Legal Agent Benchmark): In the results shared by Mistral, ML4 scored 15.8 (an absolute score, not a percentage), outpacing Kimi K3 (12.9), Qwen3.8 Max (10.4), GLM-5.3 (8.3), and GPT-6 Astra (5.4).
• Financial operations and spreadsheet handling: On the Finance Agent V2 benchmark, ML4 scored 54.7, ahead of GPT-6 Astra (53.5) and slightly behind GLM-5.3 (55.9); on FinWorkBench, which involves creating and editing real accounting spreadsheets, ML4 scored 67.4, tying DeepSeek V4 Pro (67.4) and trailing Kimi K3 (77.3). While these scores reflect model capabilities in specific tool orchestration and operational pipelines, they cannot substitute for human review by qualified legal or audit professionals.

2. Financial Statement Retrieval Demo: Cross-Year Footnote Reconciliation

In an official demonstration analyzing multi-year financial statements, the system was given a targeted query:

"What were the carrying amounts of Merck KGaA's contingent consideration assets arising from the divestment of its biosimilars business at year-end 2022 and year-end 2023? What post-tax discount rates were applied at those times? What was the year-over-year percentage change in carrying amount between the two years?"

Contingent consideration assets (contractual rights to receive future payments contingent on subsequent R&D or commercialization milestones following a divestiture) are typically buried deep within financial statement footnotes. In this demonstration, multiple evaluated models retrieved the core facts and completed the calculations, though their recorded trajectory turns varied (16 turns for ML4, 28 turns for GLM-5.3, 38 turns for DeepSeek V4, and 42 turns for MM3.5); this reflects differences in retrieval paths within this specific demo rather than a generalized advantage in time, cost, or accuracy. In this instance, ML4 reconciled footnotes across the 2023 and 2024 annual reports:

• Confirmed that the asset is classified under Level 3 Fair Value measurement (Level 3 Fair Value);
• Located the carrying amount decreasing from €219 million at year-end 2022 to €118 million at year-end 2023, calculating a year-over-year reduction of €101 million (a change of -46.12%);
• Extracted the post-tax discount rate shifting from a range at year-end 2022 (6.3%–7.3%) to a single rate of 6.6% at year-end 2023, noting based on disclosures that the reduction in carrying value was primarily driven by milestone payments received from the acquirer during the year.

3. Multimodal Understanding and Engineering Drawing Demos

Across multimodal benchmarks, different tests reveal distinct facets of capability: on the Dense 200 dense visual grounding evaluation, the original chart shows ML4 scoring 42.0%, edging out GPT-6 Astra's 41.5% (a 0.5 percentage point difference); on ChartQA Pro for chart reasoning, ML4 scored 63.1%, trailing GPT-6 Astra's 65.1%; on GDP.pdf for long-document chart Q&A, ML4 scored 18.6, below Kimi K3's 22.

In official mechanical drawing demonstrations, the model illustrates the process of reading callouts and performing assistive calculations:

• Assembly relationship analysis: Given a mechanical cross-sectional drawing and asked "Does the compression spring make direct contact with the sliding sleeve bearing?", the model zoomed in on the bill of materials (BOM) to verify part numbers, traced leader lines back to the section view, identified an intervening vacuum pin (Vacuum pin) between the two components, and concluded that they do not make direct contact.
• Tolerance callout reading and calculation: After parsing the fit specification Ø25 p6, the demo shows the model retrieving the lower deviation (+0.022mm) and upper deviation (+0.035mm) corresponding to a nominal 25mm diameter, calculating the limit dimensions as 25.022mm to 25.035mm (ISO 286 is an international standard). This demo showcases assistive capabilities pairing localized image zooming with dimensional calculations, though production engineering requires rigorous manual review against formal specifications.

4. Industrial Imagery: Pinpointing Exact Locations and Answering "Not Found"

Official visual demonstrations also showcase grounding tasks beyond technical drawings: locating blades on foreground wind turbines in a wind farm, detecting glowing red steel billets in a forge workshop using oriented bounding boxes (OBB, bounding boxes rotated along the object's orientation) for tilted targets; on printed circuit boards, locating silkscreened "GND" text, red capacitors, and transformers. Grounding results across continuous video sequences are further structured into multi-frame trajectories, allowing users to track the positional changes of a single object over time.

This output transforms "what is in the image" into "where it is in the image." Analysts can directly cross-check bounding boxes against raw frames rather than relying on descriptive text alone to infer whether the model located the correct entity.

An orange sorting conveyor demonstration highlights another valuable behavior. In the same video clip, querying for "orange" returns bounding boxes; querying for "apple" returns an empty result, with no detection frames or trajectories recorded in the original logs. The model does not arbitrarily pick a similar fruit to satisfy the prompt. Retaining an empty result when the target is absent prevents feeding false detections into downstream pipelines; this example illustrates negative-sample handling rather than a formal false-positive rate measurement on an active production line.

Query: Find oranges

Returns bounding boxes for oranges in frame

Query: Find apples

Returns empty result, with no bounding boxes

Compare frame for apple query

5. PDFs: Grounding Evidence Before Linking Charts to Body Text

A query drawn from an energy outlook report asks two things at once: in the chart, which sector contributes most to US retail electricity sales growth? In the text, how much additional solar electricity will the Texas grid (ERCOT) generate over two years? The official demo first searches for evidence across a 12-page context before reading both chart and narrative from the relevant page: the chart identifies the commercial sector as the largest contributor; the text reports additions of 18 BkWh in 2026 and 20 BkWh in 2027, totaling 38 BkWh (38 billion kilowatt-hours).

Solving this requires more than reading an isolated chart or adding numbers. It entails finding the page, parsing the legend, extracting figures for the matching years, and finally calculating 18 + 20. Another crude oil outage query follows the same workflow: reading an estimated outage of 3,550 thousand barrels/day and a baseline production of 10,500 thousand barrels/day from the same row for Saudi Arabia, calculating 3,550 / 10,500 ≈ 33.8%. The figures can be verified against the source table, and the calculation has an explicit numerator and denominator.

6. Satellite Imagery: Progressive Zoom to Locate the Rooftop "H"

The Busan Port case study requires finding a helipad located in the lower-left sector of the source image. The demo first isolates the lower-left port area, narrows its focus to perimeter buildings, and finally zooms in on a tower rooftop to locate the H marking inside a circular landing zone. A human operator can follow this identical zoom path to verify whether the identified target is correct.

The source image measures 6,144 × 4,096, roughly 25.17 million pixels; this demo image is distinct from the gigapixel imagery capabilities referenced in the general announcement. This concrete example illustrates the process of using tools to progressively narrow search areas and verify fine details.

These official demo logs reflect internal model checkpoints demonstrating verifiable coordinates and calculation steps; performance on the public preview API should still be validated on individual workloads.

Scientific Computing and Mathematical Reasoning: Code Generation and Domain Preferences

On scientific computing and mathematical modeling tasks, Mistral showcases the model's capacity for authoring scientific workflow code and quantitative reasoning:

• SciCode-Verified scientific programming benchmark: This evaluation measures the ability to write code solving complex problems across physics, mathematics, materials science, and biology. In pass@1 success rates (pass@1), ML4 reached 91.8%; among comparison models shown in the chart, Qwen3.8 2.4T A95B scored 93.8%, GLM-5.3 reached 92.5%, MiMo-V2.6-Pro stood at 91.9%, and closed-source GPT-6 Astra recorded 94.2%. In a case demonstration, the model showed single-pass generation of Hartree–Fock numerical simulation code (Hartree–Fock, a foundational quantum chemistry simulation method used to determine the electronic structure of many-electron systems); however, this serves as a code-generation demonstration—the resulting code still requires execution and validation against empirical data, and does not mean scientific conclusions can be accepted without verification.
• Domain expert evaluation preferences: In an internal expert evaluation conducted by Mistral, annotators compared and scored outputs between ML4 and GLM-5.3 head-to-head. Across different domains, weighted expert preference rates for ML4 were: STEM 68% (STEM stands for Science, Technology, Engineering, and Mathematics), Computer-Aided Design (CAD) 62%, Finance 50%, and Code 48%. These percentages reflect expert preference between model outputs rather than task accuracy rates.
• Qualitative progress in formal and applied mathematical reasoning: In quantitative logic, official narratives note that ML4 exhibits stronger capabilities in formal reasoning and applied mathematics, displaying more rigorous structural reasoning than GLM-5.3 in internal human evaluations and sustaining long-horizon, discipline-heavy applied mathematical tasks (such as derivations in cutting-edge theoretical physics).

Reinforcement Learning and Training Engineering: Scaled RL on a 3,000-GPU Cluster

ML4's pre-training was conducted on 3,800 Nvidia Grace Blackwell GPUs in Mistral's proprietary European data center, where the public preview service is currently hosted. During post-training (Post-training), Mistral deployed a dedicated cluster of approximately 3,000 GPUs to scale reinforcement learning (RL).

The large-scale reinforcement learning mechanism during the post-training phase primarily includes the following aspects:

• Large-scale rollouts and synthetic training generation: Across the 3,000-GPU RL cluster, auto-scaling actor nodes (Actors) concurrently generate tens of thousands of exploration rollouts (Rollouts, representing the model's hands-on trial-and-error trajectory in task environments), while parameter updates run asynchronously. As base model capabilities mature, conventional human annotations no longer provide sufficient challenge; the model must reinforce itself against outcomes of its own explorations, scaling task difficulty as its proficiencies grow. This pipeline generates roughly 33 billion total tokens daily, filtering and masking them to yield roughly 16 billion trainable completion tokens for policy updates (Trainable completion tokens).
• Modular composable verifier environments: To support diverse real-world tasks, the training environment provides unified composable interfaces, enabling a single RL run to encompass single-turn dialogues, complex scientific problem-solving, safety alignment, factuality checks, and long-horizon tool usage. These environments share infrastructure including code sandboxes, web search, and external APIs. For verification and feedback, reward models, unit tests, LLM judges, and static analyzers are composed dynamically based on task requirements to score rollouts and supply training signals.
• Long-trajectory policy stability: In complex tasks involving multi-turn tool calling and extended interactions, asynchronous training means the policy generating rollouts lags behind the updated policy in the optimizer (known as staleness and off-policy drift, Staleness & Off-policy drift), which threatens long-trajectory training stability. Mistral states that it adopted cross-phase optimization techniques to mitigate staleness and drift, enabling training to sustain rollout budgets of several million tokens (Rollout budget, representing the cumulative exploration token budget allowed during training after iterative context compression—a distinct concept from the 1M context window at inference).
• Phased progression vs. downstream evaluation: Mistral notes that gains inside training environments transferred to downstream evaluations, with final capabilities reflecting combined contributions from both SFT and RL stages. Across three internal downstream curves, complex coding agent metrics moved from roughly 58.5% at the end of SFT to 68.84% by the end of RL; codebase comprehension climbed from roughly 24% to 38.66% (after an initial dip to around 20% early in RL); and professional delivery tasks progressed from approximately 1,380 to 1,511. The chart does not name specific public benchmarks, meaning 1,511 cannot be equated with AA-Briefcase's 1,393 Elo, nor can 68.84% be treated as DeepSWE's 61.7%.

Rising training rewards indicate that the model is improving at tasks scored by training environment rules; downstream benchmark gains provide additional evidence that learned policies generalize to separate evaluations. Reviewing both sets of metrics side by side clarifies whether training yields genuine out-of-environment gains rather than merely over-optimizing for in-loop rewards.

Cost and Deployment Considerations: API Pricing and Evaluation Dimensions

For developers and enterprise customers evaluating production integration, clear pricing and operational considerations are essential:

1. API Pricing

Standard public preview API pricing published on the official release page is:

• Input pricing: $1.36 / million tokens
• Output pricing: $4.18 / million tokens

Developers conducting preliminary cost modeling should reference these published standard API rates.

2. Evaluation and Architecture Considerations

• Benchmark against proprietary tasks and tool failure resilience: For long-horizon workflows combining document retrieval, visual inspection, and multi-step tool calls, developers should establish domain-specific acceptance testbeds. These should specifically assess error-recovery resilience when encountering tool failures or invalid parameters, as well as final deliverable quality, to gauge real-world performance.
• Self-hosting awaits weights and license releases: For enterprises requiring strict data sovereignty and on-premises deployments, the official plan is to release model weights by month's end. Specific licensing terms, hardware requirements, and sizing guidelines have not yet been disclosed; because the model comprises 1.05 trillion total parameters, deployments must still store the entire weight footprint in memory, meaning server specifications and VRAM thresholds cannot be inferred from active 49B parameters alone.
• Avoid over-provisioning for lightweight standard tasks: For routine text summarization, basic chat, or single-file code completion, cost-efficiency should not be judged solely by parameter scale. Teams should evaluate end-to-end latency, cumulative token expenses across multi-turn tool calls, and actual output quality.

Mistral indicates that ML4 will serve as the foundation for its next generation of specialized and optimized models. Reinforcement learning training behind the preview continues actively, with further compute expansion backed by its recently closed €3 billion Series D round (the largest equity funding round for a European tech firm), bringing additional European sovereign data center capacity online. Following its plan to open weights by month's end, Mistral is articulating a distinct posture for enterprise clients: delivering frontier capabilities in mission-critical domains like cybersecurity, finance, and code, while pairing open weights with European sovereign operations so organizations can maintain control over their deployment environments and operational policies.