Many teams explore in LM Studio and stabilize in Ollama. The painful part is the handoff: a model that "worked on my laptop" becomes a vague request for IT to "put it on the server."
Make the handoff artifact explicit
Require a short package: exact filename and checksum, intended context length, system prompt file, a ten-prompt eval set, and measured resource notes from the explorer's machine. Reject handoffs that only say a marketing model name. Keep exploration folders separate from the server allowlist as discussed in operational hygiene habits alongside install tutorials.
Server acceptance
On the shared host, recreate the model via an Ollama Modelfile or pinned GGUF path. Run the same eval set. Confirm VRAM headroom with vram-table thinking, not vibes. Publish the tag name, API base URL, and owner. Point non-technical staff to a UI such as patterns in microai-one rather than raw ports.
When hardware disagrees
If the laptop was a high-memory Mac and the server is a smaller NVIDIA box, say so early using Mac versus hardware comparisons. Sometimes the right answer is a smaller quant, not a louder complaint. Troubleshooting order stays practical under troubleshooting.
Extra
Document unload behavior for night batch jobs. Keep a rollback tag. For assisted rollout see services and solutions. Community tips: community.
Handoff quality is a package, not a vibe.
Appendix: rejection template
When a handoff lacks checksums or eval prompts, reply with a template listing the five missing artifacts. Do not soft-accept "we will add later." Later rarely arrives before users complain. Point explorers at Ollama tutorials so they know what production expects.
Appendix B: communication to non-engineers
When the production tag is ready, announce the tag name, who to ping, and what changed in one short paragraph. Avoid dumping Modelfile diffs into general channels. Link a UI entry point so people are not asked to curl localhost. If the handoff fails eval, say which of the ten prompts failed without blaming individuals in public channels. Keep a rollback tag named explicitly so support can ask for it without guessing.
Also record the explorer laptop specs in the handoff so Mac versus NVIDIA surprises are visible before IT spends an evening. Require the eval outputs to be attached as files, not pasted screenshots alone.
Published by MicroAI One (本地AI 1号网), a Shenzhen team working on local LLM hardware, software and deployment.