GMKTEC EVO-X2 / ARCHITECTURE STUDIO
Your RAG stack.
One capable machine.
Choose the right specialists. See what fits. Build a self-contained, Windows-first system around the Ryzen AI Max+ 395.
RYZEN AI MAX+395
01 / CONFIGURE
Make it your stack.
Budget = weights + working memory
Saved in this browser
This is a planning tool. It does not install models, run inference, or access your documents. Your selections stay in this browser; no model API is called.
02 / PUT IT TO WORK
A complete system. A deliberate starting point.
Small specialists,
Keep the whole workflow
on the box.
- Serve + code
- LM Studio with a current llama.cpp Vulkan runtime, serving on
127.0.0.1. Connect your local editor client to that endpoint. Test generation and tool calls separately. - Retrieve + store
- Python, Qdrant local mode for a single-process vector store, plus SQLite for metadata and BM25. No Docker required for this baseline.
- Ingest + verify
- Extract PDF text first, apply local OCR for scans, and use vision for difficult pages. Preserve document revision, page and section IDs. Check citations and units before accepting an answer.
Windows compatibility & limits
Vulkan is the initial GPU path. ROCm is optional: verify the exact Radeon 8060S, Windows, driver, Python and framework combination against AMD’s support matrix. CUDA packages target NVIDIA; the NPU is outside this budget.
Run embedding and reranking on CPU via Sentence Transformers/PyTorch. Avoid assuming Linux-only acceleration dependencies work on Windows. Current versions must support IBM’s ModernBERT encoders; use small batches.
Qdrant local mode suits one local application and modest indexes. Move to a dedicated server when you need multi-process access or larger collections. PostgreSQL + pgvector is another durable option, but adds service and Windows extension setup; MariaDB is not interchangeable without changing the retrieval integration.
Local operation requires downloading software/models first and configuring local endpoints with cloud fallback disabled. This hosted configurator does not connect to the PC.
Give evidence
the advantage.
One embedder + BM25 finds both meaning and exact identifiers. Fuse ranks, rerank the best passages, then check the generated answer.
- Start with 400–800 token chunks.Respect sections. Give table rows their own searchable text and a link to the parent table.
- Retrieve 20 + 20; fuse; rerank 20.Send the best 5–8 passages to the answerer. These are starting values to evaluate.
- Measure before adding models.Use answerable and unanswerable questions. Track retrieval recall, citation support, abstention, memory and latency.
Model origin, licenses & sources
The filter means the original model developer is OpenAI, IBM, Google or Mistral AI, based in the US or France. It is not a claim about every training-data source, contributor, or quantization uploader. Chinese-developed model families and derivatives are not offered. This restriction concerns models; the hardware is the EVO-X2 you specified.
OpenAI, IBM and the listed Mistral releases use Apache 2.0. Gemma 3 uses its own terms. Follow each model card link in the role details, and verify the license and source of the exact weights you download. This is a curated practical shortlist, not a latest-model leaderboard.
Sources checked 12 September 2026. Hardware and model capabilities come from linked primary sources; budgets and architecture recommendations are design estimates. llama.cpp multimodal support ↗