TL;DR
Thinking Machines, Mistral AI and Microsoft are promoting three distinct ways for organizations to customize models and retain greater control over the results. Their offerings differ on portability, jurisdiction, infrastructure and vendor dependence, while the legal and practical meaning of model ownership remains unsettled.
Thinking Machines, Mistral AI and Microsoft are competing to sell organizations customized AI models that customers can treat as their own, moving beyond reliance on a standard hosted API. Their respective offerings—Tinker with Inkling, Mistral Forge and Microsoft Frontier Tuning—present sharply different choices over weight portability, jurisdiction and platform dependence, matters that can shape adoption in healthcare, finance and defense.
Thinking Machines’ Tinker provides a low-level training interface for models including Inkling, Qwen, DeepSeek, Kimi and other open-weight options. According to the company’s documentation as described by Thorsten Meyer AI, teams can use LoRA-based fine-tuning, control core training operations and download the resulting weights or adapters for deployment elsewhere.
Mistral Forge follows a managed-program model. Mistral works with customers across pre-training and post-training, including supervised fine-tuning and reinforcement learning. The resulting model can be deployed on premises, in European infrastructure or in an air-gapped environment, making the offer aimed mainly at data-mature regulated organizations prepared for a deeper vendor engagement.
Microsoft’s MAI models and Frontier Tuning offer weight-level customization through Azure AI Foundry. Microsoft presents this route as combining first-party model lineage with access to Foundry’s broader catalog. Customers may control their tuned model, but deployment carries strong Azure infrastructure and service dependencies, limiting the freedom to move the complete workload to another provider.
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 Choices Shape AI Risk
The differences matter most where sensitive data cannot move freely, where specialized terminology changes how a model must reason, and where procurement teams demand a clear account of model lineage. A health system may prioritize patient-data controls; a defense organization may need an air-gapped deployment; a bank may focus on auditability and long-term service continuity.
The three approaches place control at different points. Tinker favors portability and technical independence but requires experienced machine-learning staff. Forge offers managed customization and European deployment options but can create a close operational relationship with Mistral. Microsoft supplies enterprise tooling and integrated governance, while customers accept greater Azure dependence.
Ollama & Local AI: A Practical Guide to Self-Hosting, Fine-Tuning, and Deploying Open-Source LLMs for Production
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Open Weights Become a Sales Channel
The comparison follows attention around Inkling’s open weights. Thorsten Meyer AI argues that the release also serves as an entry point to Tinker’s paid customization platform: organizations can examine or adopt the base model, then use Thinking Machines’ infrastructure to train a specialized version.
The broader target is not every company. The strongest demand is expected in healthcare, finance, defense, pharmaceuticals and legal services, where a generic external API may conflict with privacy rules, classification requirements or internal risk policies. In these fields, the buying question extends beyond benchmark performance to data handling, deployment rights and model continuity.
“Inkling’s open weights were the headline; Tinker is the business.”
— Thorsten Meyer AI
LOCAL LLM DEPLOYMENT: Training, Fine-Tuning, & Offline Inference: The Complete Developer’s Guide to Building, Training, and Running Private Open-Source AI Offline (with full source code)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Ownership Claims Lack One Definition
It is not yet clear whether customers will interpret “owning” a model in the same way across the three services. Downloading a LoRA adapter, controlling a customized checkpoint and holding broad deployment rights are different technical and contractual positions. Rights may also depend on the base model’s license and the customer agreement.
The cited performance, efficiency and privacy statements are vendor claims that await independent replication. Public information also does not establish comparable pricing, support obligations, exit costs or production results across the three programs. Microsoft’s reported efficiency gains and early customer work, including references to Mayo Clinic, need fuller technical disclosure before direct comparisons can be made.
All in 1 AI Model: Official Step-by-Step Curriculum: How to Create, Launch, and Monetize AI Models – 16 Module Training Program (The Lazy Genius)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Contracts and Deployments Face the Test
Prospective customers will need to compare licenses, export rights, data-use clauses and deployment restrictions before accepting an ownership claim. The next evidence will come from production deployments showing whether customized models provide reliable domain performance without creating unacceptable costs or vendor dependence. Independent benchmarks and published customer agreements could clarify how much control each route delivers in practice.
Product Pieces: AI Solutions: The plain-English field guide to building with AI — and how every idea goes wrong.
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
Does Tinker let customers download a customized model?
According to the cited Thinking Machines documentation, customers can download trained weights or adapters. The scope of reuse still depends on the base model’s license and the applicable service agreement.
How is Mistral Forge different from Tinker?
Forge is a managed customization program covering more of the model-development lifecycle. Tinker provides a lower-level training interface that gives technical teams more direct control.
Can a Frontier Tuning model leave Azure?
The source describes the tuned model as belonging to the customer but remaining closely tied to Microsoft’s ecosystem. Exact portability depends on Microsoft’s contracts, tooling and deployment terms.
Which option offers the most independence?
On the information available, Tinker offers the strongest portability because it supports open bases and downloadable outputs. That independence requires greater internal machine-learning capacity.
Who is most likely to use these services?
The main prospective users are regulated or high-consequence organizations that need specialized reasoning, strict data controls and dependable model lineage. Many companies with ordinary workloads may still find a hosted general-purpose API simpler and less costly.
Source: Thorsten Meyer AI
