How to Get Started with Forge

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 carries over 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 form the knowledge base, and deep research is the Deep Research skill. Plans are now called Mistral Free, Pro, Team, and Enterprise. This lesson uses the name Vibe; some menus described here may have changed names or locations.

A Four-Phase Project

Launching a Forge project requires planning. It demands rigorous preparation, close collaboration with Mistral teams, and a clear vision of your objectives. Each phase sets up the next: vague scoping leads to poor data selection, poor data yields a disappointing model, and you won’t discover the problem until three months later.

Phase 1 — Assessment and Scoping

First, identify scenarios where fine-tuning delivers more value than a base model or RAG. In practice, candidates fall into four recurring categories. Start with repetitive tasks requiring deep expertise—technical diagnostics, compliance analysis, specialized code reviews—where the model must replicate expert reasoning rather than retrieve a document. Next come standardized customer interactions: tier 1–2 support, RFP responses, sector reports, where tone and structure follow your established conventions. Then come specific business processes, such as risk assessment against your internal criteria or document classification using your taxonomy. Finally, regulatory compliance applies when you must produce documents that adhere to sector-specific standards a generic model doesn’t know in detail.

Second, assess your data maturity, which directly determines model quality. Start with volume: verify you have enough high-quality examples—typically several thousand for effective fine-tuning. Then examine quality itself: are the examples reliable, up to date, and representative of your expertise? Next, check diversity: do they cover the full range of scenarios the model will handle? Finally, review format: are they already structured, or do they require preparation? The phase closes with a meeting between your team and Mistral’s commercial and technical teams: present your objectives, discuss use cases, confirm technical feasibility, estimate scope and timeline, and receive a commercial proposal.

Phase 2 — Data Preparation

Forward-Deployed Engineers build the training dataset alongside your teams. Extraction pulls data from your source systems—knowledge bases, ticket histories, technical documentation, customer exchanges. Cleaning removes duplicates, corrects errors, and anonymizes personal data. Formatting transforms everything into instruction/response pairs suited for fine-tuning. Validation hands the dataset to your subject-matter experts to verify example quality and relevance. This is the most underestimated step: an unvalidated dataset teaches the model your approximations alongside your best practices.

A separate evaluation set runs in parallel: business benchmarks—questions and answers validated by your experts—and edge cases, complex or ambiguous scenarios designed to test robustness. You ask the same questions of both the base model and the customized model, then compare them against explicit metrics: accuracy, relevance, and adherence to your internal tone and processes. Throughout this phase, data moves over encrypted channels, resides in an isolated environment dedicated to your organization, remains accessible only to Mistral engineers assigned to the project, and is covered by a specific confidentiality agreement.

Phase 3 — Model Training

Base model selection depends on the use case: Mistral Large for complex tasks requiring advanced reasoning, Mistral Medium for the best performance-to-cost balance in most situations, and specialized models optimized for code, multilingual tasks, or other domains.

Training runs on Mistral’s GPU infrastructure in Europe. Engineers configure hyperparameters, launch supervised training on your data, monitor convergence in real time, and run multiple iterations. Duration ranges from a few hours to several days, depending on data volume and base model size. After each cycle, the model is evaluated on the test set—compared against the base model, with error and edge-case analysis, data adjustments if needed, and validation by your subject-matter experts—until it meets the performance targets defined during scoping.

Phase 4 — Production Deployment

Deploy the customized model according to your infrastructure and compliance constraints.

Integration then follows several complementary paths: Mistral Enterprise, where the model becomes available in the conversational interface; the REST API, to call it from your applications, workflows, and business tools; Python and JavaScript SDKs for programmatic integration; and webhooks, which trigger the model on business events. Deployment doesn’t end the project. Continuous monitoring tracks latency, accuracy, and user satisfaction, detects drift, triggers alerts on degradation, and schedules updates and retraining.

Typical Timeline

Total duration ranges from 3 to 6 months, depending on use-case complexity and data maturity. If you start with scattered, unvalidated data, Phase 2 expands—never the training phase.

Key Takeaways

• The Forge process follows four phases: assessment, data, training, deployment
• Training data quality determines model performance
• Forward-Deployed Engineers guide every project step
• Deployment options include Mistral Cloud, VPC, on-premise, or partner cloud
• A typical project takes 3 to 6 months from assessment to production