OpenViking on a 99-Yuan 2-Core 2G Cloud Host: Renewal Math, 128K Embeddings, and a Hidden Query Planner
Self-hosted agent memory stacks don't need high-end hardware: renewal pricing, a 128K multimodal embedding model, and a dedicated query planner model decide whether a cheap cloud host works. The same cost model and configuration pattern transfer to any OpenViking deployment handling growing memory stores.
OpenViking's AI coding agent memory backend runs acceptably on a 2-core 2G Aliyun instance that renews at 99 yuan/year, a choice driven by renewal math rather than first-year promotions. All inference goes to external APIs; the machine handles storage and retrieval only, keeping memory usage at a few hundred megabytes. The central cost rule for stateful services: a host whose first-year discount balloons on renewal forces an expensive cross-cloud migration, so stable renewal pricing matters more than the initial sticker.
Embedding model selection is a one-way door. Changing it later means a full re-index of all accumulated vectors, so the choice locks in the database's geometry. The winning pick is Volcano Ark's doubao-embedding-vision-251215, a multimodal model with a 128K context window priced at 0.7 yuan per million tokens, which beat the same-priced qwen3-vl-embedding purely on that context window. Free BGE-M3 from SiliconFlow loses on recall quality and suffers from an overcrowded public free pool.
A hidden configuration switch resolves the multi-second stutter most users blame on hardware. Without an explicit query_planner, OpenViking runs every turn's intent analysis through the main VLM, which with thinking enabled can eat 6-7 seconds. The fix is a tiny domain model, tongyi-xiaomi-analysis-flash, that returns compliant JSON in about 0.8 seconds and rejects small talk in 0.4 seconds, at a price below general-purpose LLMs. Rerank stays on SiliconFlow's commercial tier, and the VLM runs at reasoning_effort: low to keep archive extraction off the critical path.
The success of a specialized model for query planning shows that intent classification is narrow enough for a fine-tuned small model to beat general-purpose LLMs on latency and format adherence, not just price.
Treating the embedding model as a one-way decision makes long-term sense because re-indexing cost grows with memory accumulation, but it also locks the stack into a single provider's geometry.
The default fallback of using the main VLM for query planning is a design trap: it couples a high-latency capability to a critical path, and the fix is a configuration detail most users never find in docs.
The renewal-price argument for cloud hosts applies specifically to stateful workloads; for stateless nodes, switching hosts is cheap and the first-year discount matters.