Mistral Forge

Mistral Forge, branded by its developer simply as Forge, is an enterprise platform and engineering service from Mistral AI for building custom large language models and related multimodal models on an organization's proprietary data. Mistral introduced it on March 17, 2026. Forge covers data preparation, pre-training, post-training, alignment, evaluation, versioning, and inference rather than offering only a self-service fine-tuning endpoint.^[1][2]

Forge is sold through Mistral's enterprise team. It is not an open-source training framework, and Mistral does not publish a Forge API reference, fixed price, complete model compatibility list, or self-service signup procedure. A project can include Mistral engineers working with the customer on data, model choice, evaluations, infrastructure, and deployment.^[2][3]

Background and positioning

Mistral offered three forms of model customization before Forge. In June 2024 it released an open fine-tuning codebase, added managed LoRA fine-tuning to its API, and offered bespoke custom-training engagements that could include continued pre-training on customer data. Forge formalized the broader engagement into a named product spanning the model lifecycle.^[4]

The distinction matters because Forge is not merely a new name for the earlier API. Its stated scope includes training a model from scratch or continuing the pre-training of a Mistral foundation model, generating synthetic data, aligning behavior, evaluating against private benchmarks, and operating the resulting model. TechCrunch reported that Forge combines tooling with forward-deployed engineers who help customers identify useful data and design evaluations.^[1][3]

Mistral announced ASML, DSO National Laboratories Singapore, Ericsson, the European Space Agency, Singapore's Home Team Science and Technology Agency (HTX), and Reply as early partners. The announcement does not say that every organization uses the same Forge configuration or that every project has reached production.^[1][3]

Training and data workflow

Forge accepts broad categories of enterprise material rather than one published file format. Mistral names internal documents, source code, structured data, operational records, ontologies, and decision frameworks. Its data-preparation stage can generate domain examples, rare edge cases, and policy-bound scenarios when the available records do not cover an important behavior.^[1][2]

The training stage supports dense and mixture-of-experts architectures. Mistral says text, image, and audio foundations can be used where relevant, making the service potentially applicable to multimodal AI as well as text language models. It has not published a universal minimum dataset size, token budget, supported media schema, or compute requirement for Forge. Those choices depend on the model and engagement.^[1][2]

Forge's public materials describe several customization methods:

• Full or continued pre-training to teach domain vocabulary and knowledge from large corpora.
• Supervised fine-tuning to shape task behavior using labeled examples.
• LoRA for parameter-efficient specialization.
• Direct Preference Optimization and reinforcement learning from human feedback to encode preferences.
• Model distillation to transfer behavior into a smaller or more efficient model.
• Reinforcement learning against internal objectives, tool use, or operational feedback.^[1][2]

These are available techniques, not a required sequence. A project may start from a Mistral checkpoint, adapt an open-weight model, or train more extensively. TechCrunch named Mistral Small 4 as one possible open-weight starting point. HTX's Phoenix-VL 1.5 Medium instead began from Mistral Medium 3.1. Mistral has not published a current matrix saying that every Mistral model supports every Forge training method.^[3][10][11]

Evaluation and lifecycle management

Forge is designed around customer-specific measures rather than only public leaderboards. The product page describes KPI-aligned evaluations, regression suites, and drift detection. Models, datasets, training runs, and configuration are versioned, with lineage and rollback to a known version when an update degrades results.^[2]

Mistral also says a code agent can use Forge in natural language to schedule jobs, search for hyperparameters, generate data, and optimize against evaluations. The launch post uses Mistral Vibe as its example. This does not eliminate the need to define correct metrics. In Mistral's interview with TechCrunch, the company acknowledged that customers often need its engineers to determine whether the data and evaluations are adequate.^[1][3]

The evaluation layer can test factual accuracy, task completion, policy compliance, latency, cost, or other internal criteria. Its effectiveness is limited by the test set: a model can improve on the selected metrics while failing on unrepresented populations, rare conditions, adversarial inputs, or changing production data. Forge does not provide a public certification that a customized model is accurate, unbiased, secure, or safe for a regulated decision.

Deployment, access, and pricing

Mistral offers inference for Forge-built models on Mistral Compute, in a customer's private cloud, or on premises. Mistral's broader services portfolio also supports managed APIs and partner clouds, but the Forge page does not promise that every custom model can be served through every channel. Model architecture, weights license, accelerators, latency target, and data controls affect the available deployment.^[2][6]

For Mistral Compute, Mistral advertises dedicated GPU clusters, managed Kubernetes or Slurm, SSO, SCIM, role-based access controls, audit trails, network isolation, AES-256 encryption at rest, bring-your-own-key support, and a data-wiping protocol. These are Compute features, not automatic guarantees for a Forge model deployed in another cloud or on customer hardware.^[7]

Forge is available by contacting Mistral. The public pricing page lists enterprise deployments and custom models as contact-sales offerings but gives no Forge training, storage, engineering, or inference rate. Costs can therefore include data work, training compute, evaluation, model optimization, deployment infrastructure, support, and ongoing retraining. Public API token prices and the older fine-tuning API's fees should not be treated as Forge pricing.^[2][12]

Mistral says customers retain control of the resulting models and can deploy without cloud lock-in. The precise ownership of weights, derivatives, training recipes, and support deliverables is not published as a universal license. Those rights, and whether model weights are delivered to the customer, depend on the contract and the license of the starting checkpoint.^[1][3]

Security, data residency, and responsible use

Forge's product page promises strict data isolation, controlled training pipelines, auditable workflows, and deployment choices intended to meet residency requirements. An on-premises deployment can keep data and inference within the customer's environment. Private-cloud and Mistral-managed installations depend on the selected region, subprocessors, network design, and commercial agreement; the Forge page does not provide a fixed worldwide residency map.^[2][6]

Mistral's general commercial data-processing addendum treats the customer as controller and Mistral as processor when Mistral handles personal data on the customer's behalf. It provides for confidentiality, security measures, breach notification, subprocessor controls, audits under conditions, and return or deletion after service termination. It is a portfolio-wide legal document rather than a Forge-specific retention schedule, so the signed agreement and deployment remain controlling.^[8]

Responsible-use mechanisms advertised for Forge include policy-bound training examples, RLHF and DPO, internal compliance evaluations, regression tests, drift monitoring, rollback, and policy-aware inference. These controls allow a customer to encode its own rules, but training on proprietary material does not establish that the material is accurate, lawful, representative, or safe. Sensitive data can also be memorized in weights, so access control and data minimization remain necessary.

Mistral's hosted-product usage policy prohibits categories such as illegal activity, fraud, privacy violations, malware, and unqualified professional advice. The policy explicitly excludes products deployed on customer or partner infrastructure. For those deployments, customer governance, applicable law, model-specific licenses, and negotiated terms are the main controls.^[9]

Open-source relationship

Forge itself has no public source repository. Mistral's separate mistral-finetune project was an Apache-2.0 LoRA trainer released in 2024. It supported local training for several Mistral checkpoints and documented JSONL data formats. Mistral archived that repository on June 16, 2026 and marked it no longer maintained.^[4][5]

The existence of that repository does not make Forge open source or reveal Forge's infrastructure and recipes. Likewise, Mistral's public fine-tuning API and its SDKs are adjacent customization tools, not documented Forge APIs. Enterprises need Mistral to specify which components, interfaces, and model licenses are part of their engagement.

Customer example

HTX provides the most detailed published Forge example. Its Phoenix-VL 1.5 Medium is a 123-billion-parameter multimodal and multilingual model developed with Mistral for Singapore public-safety contexts and deployed on HTX's own NGINE infrastructure. The accompanying technical report says the team continued pre-training Mistral Medium 3.1 on a localized one-trillion-token multimodal corpus, added a 250-billion-token long-context phase, post-trained on 22 billion tokens, and used another five billion tokens for online DPO. It also created localized knowledge, behavior, and safety evaluations.^[10][11]

This example demonstrates that Forge can support pre-training, multimodal adaptation, preference alignment, custom evaluation, and on-premises delivery. It does not show what a typical project costs or how much Forge alone contributed to the outcome. HTX supplied proprietary data, sovereign infrastructure, engineers, evaluation design, and operational requirements, while Mistral supplied models and expertise. The report's authors include both organizations, so its benchmark claims are not an independent evaluation of Forge.

Limitations

Forge is an enterprise engagement rather than a transparent, reproducible public service. Its website does not disclose a stable API, service-level agreement, hardware sizing guide, data-volume thresholds, job limits, supported-model matrix, standard contract, or rate card. Prospective customers must obtain those details from Mistral.^[1][2]

Encoding knowledge in model weights can improve domain behavior, but it is slower to update and harder to inspect than retrieval-augmented generation. Training can also cause regressions, amplify errors in internal records, memorize confidential text, or overfit private evaluations. Versioning and regression tests reduce these risks but do not remove them.

Mistral markets Forge for high-consequence areas including government, finance, manufacturing, cybersecurity, and quantitative research. In those settings, model output still requires domain validation, security testing, human oversight, and an accountable deployment process. Forge supplies customization and lifecycle tools; it does not transfer responsibility for AI safety or regulatory compliance from the organization operating the model.

References

1. Mistral AI. "Introducing Forge." March 17, 2026. https://mistral.ai/news/forge/ ↩
2. Mistral AI. "Mistral Forge: Build AI models that know your enterprise." https://mistral.ai/products/forge/ ↩
3. Anna Heim and Rebecca Bellan. "Mistral bets on 'build-your-own AI' as it takes on OpenAI, Anthropic in the enterprise." TechCrunch, March 17, 2026. https://techcrunch.com/2026/03/17/mistral-forge-nvidia-gtc-build-your-own-ai-enterprise/ ↩
4. Mistral AI. "My Tailor is Mistral." June 5, 2024. https://mistral.ai/news/customization/ ↩
5. Mistral AI. "mistralai/mistral-finetune." GitHub. Archived June 16, 2026. https://github.com/mistralai/mistral-finetune ↩
6. Mistral AI. "Mistral services." https://mistral.ai/services/ ↩
7. Mistral AI. "Mistral Compute." https://mistral.ai/products/compute/ ↩
8. Mistral AI. "Data Processing Addendum." Effective March 12, 2026. https://legal.mistral.ai/terms/data-processing-addendum ↩
9. Mistral AI. "Usage Policy." Effective June 11, 2026. https://legal.mistral.ai/terms/usage-policy ↩
10. Mistral AI. "HTX: Mission-focused AI." https://mistral.ai/customers/htx/ ↩
11. Team Phoenix and Mistral AI. "Phoenix-VL 1.5 Medium Technical Report." May 11, 2026. https://arxiv.org/abs/2605.10391 ↩
12. Mistral AI. "Pricing." https://mistral.ai/pricing/ ↩