Checked 2026-10-02. Author: Billy G.R. Retail prices move; the hardware catalog stores the Amazon snapshot, not a promise of stock.
Citeable facts
Claim. Google’s 30 September 2026 post says Gemini skills are replacing Gems. Skills were already in Gemini Spark and were rolling out to Gemini chat.
Method. blog.google post by Deven Tokuno, “Let skills in Gemini tackle your most repetitive tasks.”
As of.
Claim. The same post says Gems support will be removed starting in November for personal accounts, in March 2027 for Workspace business, enterprise, and nonprofit, and in June 2027 for Workspace education. Personal Gems migrate to skills. Gems by Google Labs do not.
Method. Quoted timeline from that post. The November line does not print a year; the post date is 30 September 2026.
As of.
Claim. A local custom agent with a saved prompt fits a 7–14B Q4 (6.2–10.4 GB) and is more reliable at 27B (18.2 GB).
Method. Site formula. Open WebUI presets and Ollama modelfiles are the local analogue, not a Gemini import.
As of.
Methodology
— parameter math, quantization bytes, and the source list.
Google is renaming the “saved custom Gemini” idea. The 30 September 2026 post Let skills in Gemini tackle your most repetitive tasks says skills replace Gems. You save instructions once and invoke them with a slash. You can stack skills and attach text, PDF, or image reference files. Skills were already in Gemini Spark and were rolling out to Gemini chat globally, with Workspace business, enterprise, nonprofit, and education customers “in the coming weeks” from that date. The productivity overview is at gemini.google/overview/productivity/. This is a Google cloud feature. It does not run on your GPU.
The dates, as the post wrote them
Support for Gems goes away on a staggered calendar. Personal accounts: starting in November. The sentence does not print “2026,” but the post is dated 30 September 2026, so that November is the next one. Workspace business, enterprise, and nonprofit: March 2027. Workspace education: June 2027. Google said existing Gems migrate into skills when Gems go away. A footnote says Gems by Google Labs turn down on the personal-account timing and do not migrate. Opal, their mini-app experiment, turns down with the personal change. I am not summarizing a help-center migration click-path I did not open. Use the help center link from the post if you need the buttons.
If your bookmark still says Gems, you are on a product Google has scheduled for removal. Do not publish a new guide titled only “Gems.” Skills is the name.
The local version of a saved skill
A skill is a prompt, optional files, and a trigger. Locally that is:
An Open WebUI model preset or a system prompt you reuse, pointed at Ollama.
An Ollama modelfile when you want the prompt baked into a local tag. The base weights still have to fit. A modelfile is not a new model size.
A LangGraph graph when the “skill” is actually a sequence of tools, not a paragraph of instructions.
You will not import a Gemini skill file and have it run. You rewrite the instructions. You also lose Gemini’s account memory and Workspace file picker. What you gain is a prompt that still works on a plane, against a model you can name.
These agents are small, until the files are not
A character-style bot does not need 70B. It needs a model that follows instructions. Formula at Q4: params × 0.6 + 2. Reference files you paste become context, which the calculator models better than a fixed table.
Local skill
Model
Memory
A tone prompt, short replies
K2-Horizon-7B
6.2 GB. A 12 GB machine is plenty.
A prompt plus a long PDF
14B Q4 or 27B Q4
10.4 GB or 18.2 GB, then KV cache for the file. 24 GB if the PDF is the point.
Several skills that call tools
One shared 27B, not a model per skill
18.2 GB once. See the CrewAI page for why copies are a trap.
Hardware for a preset, not for Gemini
Gemini Skills need no local VRAM. The local preset does. A Mac mini M5 Pro 24 GB (B0HGGHNQY6, $1,669.99 on 1 October 2026) covers the 7B and a tight 27B. I would not buy a 128 GB box to host slash-commands. Buy memory when the reference file or the model size demands it, which this use rarely does past 24 GB.
This site uses cookies and shows personalised ads via Google AdSense. We and our partners store and access information on your device to serve relevant ads and improve your experience.
You can accept all cookies, decline (non-personalised ads only), or
manage preferences.
See our Privacy Policy.
Cookie preferences
Choose which cookies you allow. Strictly necessary cookies are always active.
Strictly necessary
Session state, security, and performance. Cannot be disabled.