OpenLoc Publishes Its Source: Who Can Actually Run This TMS for Free?

On October 19 (San Francisco time), ModelFront stated on its site that the initial source code, documentation, and public roadmap for OpenLoc had been released. The GitHub repository modelfront/openloc was created at 2026-10-20 06:52 UTC, which is 23:52 on October 19 in San Francisco. The license is Business Source License 1.1, not an OSI-approved open-source license. The README also marks the project as a public beta before 1.0.

The point that needs reading first is not that the source is public. Free use is written into an additional grant, and that grant has explicit exclusions. A company using OpenLoc for its own content has a relatively wide grant. A language-service provider using it to deliver work to other companies, or someone else hosting an instance for a client, is outside the free grant. Each version converts to the Apache License 2.0 only four years after it is published.

What went public is a repository, not a drop-in replacement

The README says the system runs and that early adopters are using it, but teams should still expect bugs, missing features, and breaking changes between versions. Self-hosting comes with no warranty. Teams that want ModelFront to run it for them need to join the Early Access Program. Translation companies, consultancies, and anyone running the system for someone else need a commercial license, or the hosted service, whose price has not been published.

The roadmap updated on October 19 separates what connects today from what does not. What connects today is AI assistants over MCP, the customer API, ModelFront, OpenAI, and Claude, plus DeepL, Google Translate, and Google and Microsoft sign-in. The content-system connectors shown in the app and on the website are still placeholders, to be built after launch. A connector is listed as available only after it has run in production with a team that uses that system. The roadmap does not commit to dates.

Free use is in the additional grant, not in the press-release wording

On October 8, ModelFront announced OpenLoc on its news page and set October 19 as the date the source would become public. The announcement says teams can self-host for free, get enterprise hosting and support from ModelFront at cost, or have hosting provided by third parties, and it uses the phrase “no AI tax.” It still describes the license as one that “converts to an open-source license.” Those are product claims from the announcement. The LICENSE file in the repository is the binding text.

The additional grant allows production use for “Your Own Business Purposes,” including content in the company’s own products and services, regardless of who originally wrote that content. Translators engaged by the company may work inside an instance the company itself operates. The grant excludes four uses: providing translation or localization services to a third party, whether or not for a fee; offering the system to a third party on a hosted, managed, or embedded basis; selling translation memories or corpora produced with it; and using data produced with it to train a model that is then offered to third parties. A contractor may install an instance the company operates, but may not host it for the company. “Hosting by third parties” in the announcement cannot be read as the free grant already covering hosted operation for someone else.

What the repository documentation actually connects

The architecture notes in the repository describe a Go backend, PostgreSQL, Redis, and a React front end. The pipeline looks up the translation memory first and calls machine translation on a miss. DeepL, Google Translate, and LibreTranslate are listed as pluggable engines. Quality checks and post-editing run through the user’s own prompts. MCP lets external AI tools read and write the same project data. DOCX and PDF are handled by a Python sidecar.

That is the structure described in the repository, not a capacity or security finding from an independent audit. “No AI tax” in the announcement does not cancel the invoices of the machine-translation and large-language-model providers a team connects. The user still has to accept the result against real documents.

Gaps a Chinese project will hit now

The roadmap lists several issues as reproduced and not yet fixed. Two bear directly on Chinese: Japanese and Chinese terms are found only when they are separated by spaces, and the punctuation check reports a missing full stop on correct Chinese, Japanese, and Korean translations. The number check does not catch a dropped decimal point; the roadmap’s example is 25.16 becoming 2516. Segmentation often leaves several sentences in one segment, and the rules cannot be changed yet.

File exchange also does not yet preserve the status a migration needs. On bilingual XLIFF export, auto-approved segments are written as final, and statuses from files made in other tools are not kept on import. The read and write controls on a term-base assignment currently have no effect. Content-system connectors remain placeholders. These are known limits listed by the project itself, not an outside evaluation.

Four gates before self-hosting

• License gate: Decide whether the use is the company’s own content, or language services delivered to other companies, or hosting on a client’s behalf. The latter uses are outside the free additional grant.
• Version gate: Review it as a public beta, not as a finished system already able to replace the current TMS. Assign responsibility for backups, upgrades, and breaking changes.
• Chinese gate: Import a real termbase, Chinese terms with no spaces, and a document that contains decimals. If term lookup, number checks, or segmentation fail, do not move production work.
• Migration gate: Check translation memory, terminology, segment status, and content-system connections separately. Downloadable source does not mean the current workflows and connectors can come along.

OpenLoc has put the source of a translation management system into a public repository, so a company can judge self-hosting from the license text rather than from the words “open-source TMS.” For teams selling abroad, self-hosting does lower the threshold for discussing platform lock-in when the scope is only the company’s own product content. It does not remove responsibility for terminology, quality, and delivery, and it does not make production use by a language-service provider free. Websites and software that keep changing still need a separate check of whether resource files and terminology can move out of the existing website and software localization workflow.

Sources: OpenLoc’s statement that the source was released on October 19, 2026; the Business Source License 1.1 text in the OpenLoc repository; the roadmap in the same repository, updated October 19, 2026; ModelFront’s October 8, 2026 announcement. A summary of the license is not a substitute for the LICENSE file. Features, known defects, and hosted pricing can still change.