Self-hosted coding models need a data-control case

A self-hosted coding assistant deserves consideration when control over source code is a requirement the available hosted service cannot meet. Lower inference cost may follow in some workloads, but it is a weak starting assumption. Running a model transfers responsibility for capacity, updates and access control to the team operating it.

DeepSeek-Coder-V2 made that decision more interesting in June 2024. Its paper, submitted on 17 June, reported coding results competitive with prominent hosted models. Those were the authors' benchmark results, not proof that the model would handle a particular company's repository well. Looking back from November 2026, the useful lesson is the decision method, rather than a recommendation to adopt that historical model today.

Downloadable weights are only the first permission

The DeepSeek-Coder-V2 repository distinguishes its MIT-licensed code from the separate licence governing the model. It states that the model series supports commercial use. The distinction matters: seeing an MIT badge on a repository does not establish the terms for the weights inside a deployed service.

The model licence includes use restrictions and conditions on distribution. Its displayed version date is 23 October 2023; that is the licence's date, not the launch date of Coder-V2. Commercial permission is therefore conditional, not equivalent to an unrestricted grant for every application.

The contrast with Codestral's 29 May 2024 announcement is instructive. Mistral described the downloadable model as available under its Non-Production License, with commercial licences available separately. Its explanation of that licence distinguishes experimentation from production use. An engineering team should not assume that an assistant used inside a revenue-producing business is outside the licence question merely because customers never see its interface.

These are historical licensing observations, not a legal determination for a deployment. Check the exact model artifact, licence revision and intended use before adoption. Later models carrying similar brand names may have different terms. A current product catalogue cannot silently replace the terms of the 2024 weights being discussed.

Describe where the code can go

Start with an illustrative client repository whose contract forbids disclosure to an unapproved external service. The immediate choice is to keep that repository out of the service. A self-hosted model is one possible route forward if the full deployment is permitted. Another is to seek the necessary approval or do the task without an AI assistant.

The model process is only one part of the data path. Review what the editor extension sends, where repository indexing runs and what prompts or generated responses are logged. Include crash reports and backups. A model running on your own server does not establish that the surrounding tools never communicate externally.

Nor does renting a GPU under your account automatically resolve a restriction on third-party access. The infrastructure provider, administrators and support arrangements may matter to the contract. Determine the allowed boundary with the people responsible for that agreement, then design within it. Self-hosting is a means of implementing a boundary, not evidence that every confidentiality obligation has been satisfied.

Own the operating workload before claiming savings

The repository offers models at different sizes and describes a BF16 inference configuration for the full model requiring eight 80 GB GPUs. That is a stated configuration for the full model, not a minimum for every variant, quantization or workload. Active parameter counts in a mixture-of-experts model must not be mistaken for the total weight memory that deployment requires.

The practical estimate should include the chosen model representation, context length and expected simultaneous users. A machine that loads the weights is not necessarily a service that responds acceptably when several developers are working. Queueing, capacity headroom and recovery after a failed update belong in the estimate.

Someone must also own the software around inference: authentication, patching and rollback. An assistant can be useful without being allowed to execute commands. If execution is added, define the repository access and credentials separately. Running inference locally does not make generated commands trustworthy.

No model was deployed or benchmarked for this article. These are evaluation questions, not a tested installation recipe. A hardware shopping list would be misleading without a selected artifact and a measured workload.

The hosted alternative has a serious case

A hosted service can be the more responsible choice when its terms and data handling fit the work. A provider may operate capacity and updates more effectively than a small team could. Developers may also complete tasks with less review effort using a model that performs better on their specific code.

Compare permitted alternatives on representative tasks. Include review time, unsuccessful attempts and the work needed to maintain the service. A low price per generated token does not establish a low cost per accepted change. Equally, a large hosted model's benchmark lead does not erase a contractual restriction.

Use non-sensitive or appropriately approved examples for the initial comparison. Define success before inspecting answers: the intended behaviour, unacceptable edits and required checks. Keep the same task and tooling conditions where possible. The result may justify a smaller local model for one job and no assistant for another.

The decision order should be clear: determine the allowed data boundary, confirm the rights to use the chosen model, then assess whether the operating cost and task quality justify it. That is also a useful starting brief for an AI agent project. The strongest reason to self-host is a control requirement you can explain and enforce; cost savings remain something to establish afterwards.