Code assistance has now become part of everyday life for many development teams. However, most of these tools send source code to cloud services. For projects with sensitive data, strict confidentiality agreements or regulatory requirements, this is a recurring problem. A new model called Flint addresses exactly this point.
What Recursive has presented
Flint is backed by the AI start-up Recursive, founded by former Deepmind researchers. It is the company's first language model: an MoE model with 40 billion parameters, designed for locally executed coding tasks.
MoE stands for Mixture of Experts – an architecture in which a model consists of several specialised subnetworks and only a portion of them is activated per request. The practical effect: a model can have many parameters without the entire computational load being incurred for every single request. It is precisely this property that makes such architectures interesting for operation on your own hardware.
Created with AI support
What is remarkable is less the model itself than the way it came about. Recursive states that it developed Flint within four months – with the help of its own AI system called "Eureka Machine". This system is said to independently test and optimise architecture variants, thus taking over parts of the research work that would otherwise be carried out manually.
The company is thus pursuing a thesis that has long been discussed in the industry: that AI systems can accelerate the development of subsequent AI systems. Whether this actually results in a self-improving cycle remains open for now. Flint is a first result, not proof of the far-reaching claim. This classification should be kept in mind before drawing conclusions about the maturity of a research approach from a single model release.
Why local models are gaining relevance
For small and medium-sized enterprises, the question of where AI models are executed is not a technical side issue. It determines which projects are permitted to work with code assistance at all.
- Data sovereignty: Source code, configurations and business logic do not leave your own infrastructure.
- Contractual commitments: Customer projects with confidentiality agreements can be represented more cleanly when no third-party providers are in the processing path.
- Predictable costs: Instead of usage-based billing, primarily hardware and operating costs are incurred.
- Independence: No dependence on price changes, model discontinuations or availability disruptions of an external provider.
These are offset by real efforts. Local models require suitable hardware, someone who is responsible for operation, and integration into the existing development environment. In addition, model versions must be maintained – a topic that the providers take care of in the cloud.
Specialisation instead of a universal model
Flint is explicitly designed for coding tasks. This focus corresponds to a development that has been emerging for some time: instead of a single large general-purpose model, companies are increasingly using several specialised models, each covering a clearly defined task.
In practice, this is good news. A model specialised in code does not have to cover the same scope as a universal assistant and can therefore be operated more readily on manageable hardware. What matters then is not absolute performance compared with the largest cloud models, but whether the quality is sufficient for the specific tasks in the team – refactoring, test creation, explaining unfamiliar codebases.
How to assess this
Announcements of new models are currently appearing at short intervals. For project managers, it is of little help to evaluate each individual one. It is more sensible to clarify the framework conditions in your own organisation before a specific model is selected.
- Which projects may not use cloud services for data protection or contractual reasons? That is where the actual need arises.
- Which tasks should the assistance actually take on? The narrower the focus, the more likely a smaller model will suffice.
- Which hardware is available or can be procured, and who is responsible for operation?
- How is quality verified? Without your own test basis of real tasks, any model choice remains speculation.
Once these points have been clarified, new releases can be assessed quickly – and there is no need to reassess with every announcement.
Assessment for web and software projects
For web and software projects, a fundamental question thus shifts: code assistance is no longer automatically tied to a cloud service, which opens up new options for projects with strict data protection requirements. At the same time, local operation creates additional responsibility for hardware, integration and quality assurance that must be taken into account in project planning. It therefore makes sense to define data protection requirements and usage scenarios early on, instead of leaving the decision about the tool to whichever model announcement is current at the time.