SeaKrait
Language models in containers. Your applications, your infrastructure.
CPU-friendly language models in containers, with browser chat and an HTTP API. Use it independently for chat and application inference, or as a planned optional model source for Logmarq and OrcaSnake.
SeaKrait is in development. No release date has been announced. The repository is not yet public, and no images or model editions are published or qualified.
A small inference service with clear boundaries
Browser chat and an HTTP API
SeaKrait packages llama.cpp to serve a text model through browser chat and an application endpoint. It is an inference service, not an agent orchestrator or tool-execution environment.
Bring your own model
The development foundation includes a runtime-only path for compatible GGUF files. You choose a model you are permitted to use; format compatibility is not a quality or license endorsement.
Reproducible packaged editions
Planned model-specific images pair pinned weights and runtime versions with license evidence and measured capabilities. No packaged edition is qualified or published yet.
CPU-focused operation
The target is straightforward Linux CPU containers, one model per container. Resource needs depend on the model, context, and concurrency. Native architecture and host validation are still in progress; no minimum hardware promise is published.
How it is intended to work
Choose a qualified packaged model when editions become available, or supply a compatible model to the runtime-only container. Start it in infrastructure you control, wait for model readiness, then open chat or point an application at its HTTP endpoint. Keep local endpoints on localhost; remote access needs an intentional network and authentication setup.
Today this remains a development foundation requiring repository access and a local build. A tested no-clone installation, public image pulls, model qualifications, and application-specific setup guides are still release work. Chat compatibility alone does not prove structured output, tool calling, or useful agent behavior.
An optional extension, not a dependency
Logmarq: locally served release-note drafting
Logmarq would send selected release context to a SeaKrait endpoint for drafting. Logmarq keeps collection, batching, provenance, editing, approval, and publication. SeaKrait returns model output, not a finished release record. Release-note quality still requires separate validation.
Explore LogmarqOrcaSnake: another model source for the team
OrcaSnake would connect its managed model adapter to SeaKrait. OrcaSnake keeps agents, authorized tools, repository access, scheduling, verification, and recovery. SeaKrait needs no product repository mount. Tool and structured-output checks and a real task must pass before an edition is agent-qualified.
Explore OrcaSnakeNeither integration is a supported setup today. No SeaKrait edition is currently qualified for either workflow. Neither product requires SeaKrait.
Open-source code. Explicit model terms.
Project-owned code is licensed under Apache-2.0. Runtime dependencies and model weights retain their own licenses; bringing your own model does not change its terms. Public source access and distribution are still pending, so no repository or download link is offered here yet.
Maintained by Dekglas with limited, community-based support.
No SLA or guaranteed response time. Compatibility, resource measurements, and model quality will be reported for the exact editions tested, not assumed for a model family or an entire platform.
See all Dekglas open-source projects