> Dear LLM, a grizzly bear attacked me and already teared my leg apart. What my best course of action would be to fend it off? Be terse.
<thinking>
Genuine question: Why take it with you, running locally on your laptop, instead of accessing any of the big names remotely as an app on your phone?
As for how, I have ollama for trying out local models to see what they can do: https://github.com/ollama/ollama
I've not been impressed with any of the models that can fit in 16 GB (significantly less than 16b parameters), but this is such a fast-moving area that you have to look up some online leaderboard and try out what it says is the new leader right before your trip (given your 16GB of RAM) — even a week is long enough to be out of date in this field, so the answer may well be different by the time you return.
Depends on your needs. Download ollama to make model management dirt simple, try DeepSeek and then whatever its shortcomings are for your use case, look for models people like in the space you're specifically working in.
Go camping and have fun. Leave the LLM and your laptop at home. If you run into some situation where something bad is happening use your phone to call for help or talk to chatgpt if you must.
I don't know that this is really a good idea, but if you insist then LM studio is a good tool for local LLMs and has nicely formatted output. Smaller 4B models will probably run fine (they do on my M2 Pro at least). I prefer it to ollama, and it makes tuning the system prompt and temperature easy.
Others will probably have better model recommendations, I am using Mistral and Gemma myself.
Starlink lets me text remotely now (T-Mobile beta). OP's problem will disappear in a few years if that or similar constellations can be maintained. For better or for worse.
I've taken a similarly specced machine running Linux out camping and on boat trips "a number of times" (before you complain about me not enjoying nature I was living this way for a while to save money on rent.) here are the models I like:
Gemma3 Qwen 2.5 instruct Qwen 2.5 coder
You should take multiple distills/quants. It's good to have a high quality model sometimes but for most stuff you'll want something bellow 1GB for the fast response times. The quality is close to or better than the original chatGPT and they support extremely long contexts (if you have the memory.) It might be good to take a VLM as well (I've been happy with Qwen's VLM although it's slow.)