TL;DR
Thinking Machines’ release of Inkling’s open weights has focused attention on Tinker, its platform for training portable model adapters. Tinker, Mistral Forge and Microsoft Frontier Tuning now present enterprises with different trade-offs between independence, jurisdictional control and managed integration.
Thinking Machines’ release of Inkling’s open weights has brought its Tinker training platform into sharper competition with Mistral Forge and Microsoft Frontier Tuning, giving regulated enterprises three distinct routes to models customized with their own data and expertise. The choice matters because each platform offers a different balance of weight portability, operational support and ecosystem dependence.
Tinker provides a low-level training interface while Thinking Machines operates the underlying computing infrastructure. Its four core functions cover gradient computation, optimizer steps, sampling and state saving, allowing experienced machine-learning teams to manage much of the training process without running their own GPU clusters.
The platform uses low-rank adaptation, or LoRA, which trains smaller adapters instead of modifying every parameter in a base model. Thinking Machines says this method can match full fine-tuning for many tasks while using less computing capacity. Tinker supports Inkling and third-party open models, including models from Qwen, DeepSeek, Kimi and Nemotron, and allows customers to download trained checkpoints.
Mistral Forge takes a more managed approach, covering pre-training and post-training through supervised fine-tuning and reinforcement learning. Mistral markets Forge to organizations seeking on-premises, European or air-gapped deployments. Microsoft offers another model through MAI and Frontier Tuning in Azure AI Foundry, combining first-party models, weight-level customization and integration with Microsoft’s cloud services. Microsoft says customers own their tuned models, though deployment remains closely tied to Azure.
Three ways to own your model: Tinker vs Forge vs Frontier Tuning
Inkling’s open weights were the headline; Tinker is the business. Three serious players now sell the same promise to the same buyer — a model that’s yours, not a rented API — in three different ways. For health, finance & defense, the differences are the whole decision.
For the regulated, defense or health buyer it reduces to one question: what do you most need to control — the weights, the jurisdiction, or the integration? None is strictly best; they’re bets on what you value. The meta-signal: three of the most sophisticated players independently concluded the future enterprise product isn’t a model you rent — it’s one you own and adapt, with your institutional knowledge as the moat. Tinker = portability & open base · Forge = depth & EU sovereignty · Microsoft = lineage & integration. The only wrong move left is renting a generic model and hoping.
Control Comes With Different Trade-Offs
The platforms are aimed most directly at healthcare, banking, defense, pharmaceuticals and legal services, where sensitive data, specialized terminology and formal procurement reviews can limit the use of generic hosted models. Buyers may need to document training-data lineage, model ownership and deployment location before placing an AI system into production.
Tinker offers the greatest apparent portability and reversibility, but it expects customers to supply experienced researchers and engineers. Forge offers deeper vendor involvement and a stronger European sovereignty position, while increasing dependence on a long-running Mistral engagement. Microsoft supplies extensive infrastructure and enterprise integration, but customers may face Azure ecosystem lock-in. No option is best for every buyer; the decision turns on whether an organization values independent weights, jurisdictional control or operational support most.
Rust for AI and Machine Learning: Build Faster, Safer, High-Performance Models with Practical Techniques for Training, Inference, and Deployment
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Enterprise AI Moves Beyond Rental
Many companies currently access advanced AI through hosted application programming interfaces, paying for usage while the provider controls the base model, deployment environment and service terms. That arrangement can shorten development time, but it may leave customers exposed to model deprecation, pricing changes and data-governance concerns.
The three platforms reflect a wider move toward enterprise-owned or enterprise-adapted models. Tinker starts with open bases and downloadable adapters; Forge offers a vendor-led development program built around Mistral checkpoints; Microsoft combines customized models with Foundry’s broader catalog. The common commercial premise is that institutional data and domain knowledge can produce more useful systems than an unchanged general-purpose model.
“Inkling’s open weights were the headline; Tinker is the business.”
— Thorsten Meyer AI Dispatch
LM Studio for Beginners: Run Private AI Models on Your Own Computer — No Cloud, No Code, No Subscription
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Performance and Ownership Claims Need Testing
Public information does not yet establish how the three offerings compare on total cost, training reliability, security controls or production performance under equivalent workloads. Claims about LoRA efficiency, ownership protections and deployment flexibility come mainly from the vendors and await independent replication. It is also unclear how easily customers can move trained systems, evaluation pipelines and operational tooling between providers, even when they can export the weights.
Fine-tuning Large Language Models Handbook: Customize GPT and Open-Source LLMs for Specialized AI Applications, Domain Adaptation, and Enterprise Solutions
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Procurement Tests Will Settle the Contest
Enterprise buyers are likely to seek technical trials, contractual guarantees and security reviews before selecting a platform. The next evidence to watch will include independent benchmarks, published customer deployments and clearer terms governing checkpoint export, data retention and model ownership. Those details will show whether Tinker’s portability, Forge’s managed sovereignty or Microsoft’s integration proves most persuasive in regulated production environments.
AI Engineering and Agentic AI: Designing Autonomous Language Model Systems with Memory, Tools, and Safe Deployment
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Tinker?
Tinker is a training API from Thinking Machines that lets customers fine-tune supported open models while the company operates the computing infrastructure. Customers can control training logic and download their trained checkpoints.
How does Mistral Forge differ from Tinker?
Mistral Forge is a managed program covering more of the model-development lifecycle, including pre-training and post-training. Tinker gives technical teams more direct control, while Forge emphasizes vendor support, European deployment and sovereign infrastructure.
Does Microsoft Frontier Tuning provide model ownership?
Microsoft says the tuned model belongs to the customer. The offering remains integrated with Azure AI Foundry, however, so practical portability may be lower than with downloadable open-model checkpoints.
Which option offers the most independence?
Based on the available descriptions, Tinker offers the strongest portability because customers can select from several open bases and export trained weights. That independence comes with a need for greater in-house machine-learning expertise.
Are vendor performance claims independently confirmed?
No equivalent independent comparison is available in the supplied material. Efficiency, security and performance statements should be treated as vendor claims pending outside testing.
Source: Thorsten Meyer AI
