---

ollama

Install instruction

Output:

https://docs.ollama.com/linux

Launch environment

#!/bin/bash

# maximum number of parallel requests per model
export OLLAMA_NUM_PARALLEL=1
# extra debug output
export OLLAMA_DEBUG=1
# Keep models loaded for a day
export OLLAMA_KEEP_ALIVE=36000
# Bind on all interface make ollama available on local network
export OLLAMA_HOST=0.0.0.0

ollama serve

First run

ollama pull 

Access ollama on remote machine

OLLAMA_HOST=http://192.168.0.55:11434 ollama run

Create new model

Tip

Output:

Over-and-over again, when a new model come out, I made a test run with chat, code
generating, text review, tool usage, etc, pick up the most capable model for these task,
scale the context size to mostly fit into the GPU (max-1 GB to run with other embedding models in parallel,
as some application loads them). Then use the model name and context length to create a model called default:latest.
All my applications VSCodium/KiloCode, Khoj, Obsidian, LangChain, Fabric, all of them using
default:latest, which can be replaced when a new better, shinyer model has come and no need the
hustle to update model name and context length everywhere.

Scale up model to GPU size

# Run ollama with custom context length
OLLAMA_CONTEXT_LENGTH=164000 ollama serve

# Run your model (load at least)
ollama run qwen3.5:9b

# See the GPU usage
ollama ps

Output:

NAME          ID              SIZE     PROCESSOR    CONTEXT    UNTIL              
qwen3.5:9b    6488c96fa5fa    11 GB    100% GPU     164000     4 minutes from now

Dump model and save to file

ollama show --modelfile qwen3.5:9b > default.Modelfile

Edit the modelfile

edit default.Modelfile

Output:

Change "FROM " and "PARAMETER num_ctx 164000".
I also set "PARAMETER temperature 0" this makes model less creative
  but more deterministic.

Delete old model

ollama rm default:latest

Create new model

ollama create default:latest -f default.Modelfile