Running models locally with Ollama
I have explored using Ollama to run models locally for various applications, including game development and automated agent workflows.
A topic page: a summary of what my posts say, with each claim quoted from the post it came from. The posts are the authority. Last checked 2026-11-06.
Summary. I have explored using Ollama to run models locally for various applications, including game development and automated agent workflows. These experiments show that local models can produce high-quality results when the tasks are well-defined and the prompts provide sufficient detail.
What the posts say
A local 30B model can generate a fully functional game without any external data transmission or API costs.
“Ollama added a
launch claudecommand that starts Claude Code backed by a local model instead of the Anthropic API.” — Claude Code (Part 11)
The local model produced a working game that matches the specifications.
“The screen matched the specification: five rows of aliens, the red UFO crossing the top, arched green shields taking damage, the score and wave counter, and the lives shown as small cannons.” — DeepSeek Harness (Part 2)
The local model produced the exact same code as the reference build except for minor formatting.
“Apart from blank lines and final newlines, they were the same, line for line, in all 14 files.” — DeepSeek Harness (Part 2)
The local model required more detail because a specification for a cloud model is too loose for a 30-billion-parameter model.
“A specification good enough for a frontier cloud model can be far too loose for a 30-billion-parameter local model.” — DeepSeek Harness (Part 2)
Local models are reliable only when the job is small and has one thing to get right.
“The local model is reliable when a job is small and has one thing to get right.” — DeepSeek Harness (Part 3)
The local model only saves money when the code it writes is much longer than its description.
“The local model saves money only when the code it writes is much longer than its description.” — DeepSeek Harness (Part 3)
The 30B model required over 41 GB of memory to run.
“Ollama was using 41.48 GB of memory to run the 30B model — essentially the whole machine.” — Claude Code (Part 11)
The local model successfully processed approximately 2 million tokens with an 81% cache hit rate.
“The local model took just over an hour for the same game.” — DeepSeek Harness (Part 2)
The posts behind this page
- Local Agent (Part 2) — How the blog assistant works now: a gatekeeper in front of a hosted Qwen3.5 9B, an agent loop that runs in the browser, and a checked wiki of the posts
- DeepSeek Harness (Part 3) — A hosted DeepSeek V4 Pro model builds Space Invaders and hands 15 small functions to qwen3-coder:30b subagents running on my Mac: every piece written locally, a playable game in 25 minutes for $0.40, and the two limits I found along the way
- DeepSeek Harness (Part 2) — Checking whether DeepSeek Harness's everything-is-a-plugin claim extends to the model itself: pointing it at a local Ollama-served model through a custom OpenAI-compatible provider, with no DeepSeek API key at all
- Agent Orchestration (Part 1) — Comparing three LangChain multi-agent patterns, subagents, skills and handoffs, on eleven conversations that need a clarifying question or a follow-up, plus a reproducible tool-call crash traced to one specific structural cause
- Agent Orchestration (Part 3) — Finding a real agent through a2a-registry.org, then orchestrating it from a LangGraph graph with a local Ollama model: discover the Agent Card, delegate over A2A, verify the reply against the official source, and pause for human approval
- Model Context Protocol (Part 3) — Building a Retrieval-Augmented Generation stack and finally exposing the search function to Claude Code over MCP
- Agentic AI for Professionals (Part 2) — Building the RAG core of a legal research assistant entirely on a 32GB M4 MacBook Air — page-safe chunking, a provider-agnostic LLM layer, and citations that can't be hallucinated
- Local Agent (Part 1) — A conversational AI assistant for this blog using WebLLM (in-browser) and a self-hosted Ollama server (Hosted, public, HTTPS) as interchangeable backends
- LLM Wiki — Building a Claude Code-powered personal knowledge base inspired by Andrej Karpathy's LLM Wiki — ingesting PDFs, emails, and notes into structured wiki pages
- Claude Code (Part 11) — Running Claude Code against a local Ollama model (qwen3-coder:30b) instead of the Anthropic API — no API costs and no data leaving the machine