The Forge Process

Update November 2, 2026 — Le Chat is now Vibe. As of May 28, 2026, Le Chat has been rebranded as Vibe (still at chat.mistral.ai): a single agent for work and code that retains your conversations, settings, and subscriptions. The legacy Le Chat experience remains available in Chat mode. In Work mode, agents are now skills (invoked with @), memories serve as the knowledge base, and deep research is the Deep Research skill. The plans are now called Mistral Free, Pro, Team, and Enterprise. This lesson uses the name Vibe; note that some menus described here may have changed names or locations.

From evaluation to deployment: the complete journey

A Forge project follows four phases, guided by a Mistral Field Deployment Engineer (FDE). You do not configure anything alone from a console. Instead, the project is a joint commitment between your organization and Mistral AI, with defined milestones, measurable deliverables, and a predictable timeline: 2 to 4 weeks for evaluation, 4 to 8 weeks for data preparation, 4 to 6 weeks for training and iteration, and 2 to 4 weeks for deployment and monitoring. In total, expect 3 to 6 months depending on the complexity of the use case and data quality. Knowing this sequence upfront prevents you from promising your leadership a result in six weeks.

Phase 1: Evaluation (2 to 4 weeks)

The goal is to determine whether Forge is the right approach and to define success criteria. The Mistral team works with your business sponsor and technical team on four fronts. The use case audit starts with the simplest yet most uncomfortable questions: What exact problem are you solving? Have you truly tested prompt engineering and RAG? Why are they insufficient? The data assessment measures available volume and its actual state—structured or unstructured, annotated or raw, current or outdated.

Next comes baseline measurement, a step many skip and later regret. You must evaluate the general-purpose model’s performance on target tasks; without this, you cannot quantify what Forge adds. The phase concludes by defining KPIs: precision, recall, processing time, and user satisfaction.

The deliverable is a scoping document containing the use case definition, data inventory, baseline metrics, target KPIs, and a provisional schedule. A joint go/no-go decision follows. If data is insufficient or the use case does not justify Forge, the Mistral team will steer you toward RAG or prompt engineering instead. A “no” at this stage saves you several months of work and a significant budget.

Phase 2: Data preparation (4 to 8 weeks)

This phase builds the training dataset. It is the most critical and time-consuming stage because model quality depends directly on the quality of the data used to train it. Work begins with collection: extracting documents, examples, and use cases from your internal systems. Next comes cleaning—removing duplicates, outdated documents, and corrupted data—followed by structuring, which transforms raw data into instruction-response pairs usable for training. When the use case requires it, annotation by your subject matter experts adds labels, categories, or metadata. A final validation by these same experts ensures the corpus reflects real-world practices and the expected quality level.

Document role allocation in writing from the start. Your data team handles collection, cleaning, and structuring. Your subject matter experts handle annotation and validation. The Mistral FDE provides methodological guidance, validates formats, and recommends volumes and quality standards.

Four pitfalls recur constantly. A corpus that overrepresents certain cases produces a model biased in the same proportions. Outdated documents create a model that confidently answers based on obsolete procedures, which is more dangerous than no answer at all. A few hundred examples are never enough for quality fine-tuning. Finally, if you must make a trade-off, choose a smaller, homogeneous corpus over a massive one of uneven quality.

Phase 3: Training and iteration (4 to 6 weeks)

The base model is fine-tuned on your dataset, then evaluated on a validation set composed of data it never saw during training. Results are compared against the baseline and target KPIs defined in Phase 1. Any gap triggers an iteration: adjusting hyperparameters, enriching the dataset, or correcting identified errors. In parallel, your experts conduct business validation by testing the model against real cases. A model that meets metrics but fails on everyday files is not ready.

In practice, reaching the target level requires 2 to 4 iterations, each lasting 1 to 2 weeks. The FDE owns the entire technical side: configuring training, optimizing hyperparameters, diagnosing performance issues, and recommending data improvements. Your team needs no machine learning expertise; the FDE acts as your technical interface with the Forge platform.

Phase 4: Deployment and monitoring (2 to 4 weeks)

The custom model deploys to the target infrastructure—Mistral Cloud, a dedicated VPC, or on-premise—and integrates into your environment via Mistral Enterprise, Enterprise Search, or your internal applications through APIs. Load tests validate performance under production conditions. Training prepares end users for a model that behaves differently than a general-purpose one. Dashboards track performance, volume, and satisfaction.

This monitoring directly informs governance. It detects gradual performance degradation, unanticipated use cases, and signals calling for retraining, such as new data availability or procedural changes. A Forge model is not static. Plan a quarterly cycle for rapidly evolving domains, a semi-annual cycle for stable domains, and on-demand cycles for major changes like new regulations, mergers, or product launches.

Budget and resources to provision

On the Mistral side, the commitment includes a dedicated FDE for the project duration, training infrastructure (GPUs), and technical and methodological support. On your side, reserve 10% to 20% of your business sponsor’s time, a data team of one to two people part-time during Phase 2, a few weekly hours from subject matter experts for validation, and a project budget that varies by complexity (to be scoped during the evaluation phase). Actually provision these resources. A project where the sponsor has not cleared this time will stall in Phase 2.

Key takeaways

• The Forge process takes 3 to 6 months across four phases: evaluation, data, training, and deployment
• Data preparation is the most critical phase—data quality determines model quality
• A Mistral FDE guides you end-to-end—your team needs no machine learning expertise
• Plan a retraining cycle to maintain performance over time