VM302 is a Windows 11 development workstation dedicated to polyglot development and the evaluation of programming assistants. The machine has no local GPU: it consumes a separately hosted inference capacity, which separates the client workstation from the compute service.
✅ Validated base & toolchains
- Virtualized Windows 11 base, without a local GPU accelerator
- VSCodium, Git and PowerShell 7 validated through functional tests
- Guest identity, network, guest agent and peripherals re-validated
- Python 3.13 — package management, virtual environment, offline test of standard libraries
- Node.js 24 LTS and npm — running a local JavaScript test
- DeepSeek Harness in a pinned preliminary version: lockfile, package integrity, dependency audit, install scripts and native modules checked
- Harness interface limited to the local machine, first profile kept read-only
🔎 Remote inference & persistence
- First validated remote inference with
qwen3.6:35b: the control prompt returned exactlyDSH_VM302_QWEN36_OK, with no tool call - After a reboot of the machine: versions, configuration and session history persisted
- New session confirming the expected response and remote GPU acceleration
- Access to the inference service limited to the private network, with no direct publication of its API on the Internet
- Automation services checked at the health level, without yet declaring consumption by an agent
🧭 Architecture — client workstation, separate compute
separation between the dev workstation and the inference service
VM302 (Windows 11, no local GPU)
├── VSCodium · Git · PowerShell 7
├── Python 3.13 · Node.js 24 LTS
└── DeepSeek Harness (local interface)
└── separate inference service (private network)
└── remote GPU accelerationThe workstation stays lightweight and GPU-less: compute is delegated to a separately hosted inference service, reachable only on the private network. This separation isolates the development tooling from the model service.
🎯 Demonstrated skills
- Preparation and validation of a virtualized Windows workstation
- Design of a polyglot development toolchain
- Controlled installation of an agentic tool based on a Cordis plugin architecture
- Integration of a GPU-less workstation with a separate inference service
- Selection of free components and separation of proprietary dependencies
- Reproducible tests and offline validation
- Separation between client workstation, inference service and automation
🚧 Limits & status:
- the Rust and C/C++ toolchains remain future work;
- the comparison of assistants and the execution of an automation workflow by an agent are still to be done;
- DeepSeek Harness is still shipped as a preliminary version: it is not presented as a stabilized component;
- the Windows license and proprietary hardware components prevent presenting the whole as “100% open source”.
Agents IA (VM301)
Outils libres & DevOps