Troubleshooting local LLMs and keeping up
The catalog of common local-LLM failures: OOM, slow tok/s, garbage output, instruction drift, RAG miss, tool-call hallucination. Plus where to follow the field as it moves.
Blog
#llm
Posts on AI engineering, LLM systems, and software development.
Fine-tuning a model locally
When fine-tuning is the right answer (rarely) and how to do it on consumer hardware: LoRA, QLoRA, MLX-LM, Unsloth. A worked example fine-tuning Llama 3.2 3B on a 16GB Mac.
Local agents and tool use
Function calling on open models in 2026: which models actually work (Qwen 2.5, Hermes 3, Llama 14B+), why local agents fail when they fail, and how to build defensive scaffolding around them.
Local RAG and embeddings
A complete local RAG pipeline in 30 lines: nomic-embed-text for embeddings, Chroma for the vector DB, Llama 3.2 for the chat model. Why local RAG often beats cloud RAG for personal knowledge bases.
Integrating a local LLM into your workflow
Wire your local LLM into VS Code (Continue, Cline), web UIs (Open WebUI, LibreChat, Page Assist), and your own apps via the OpenAI-compatible API. The swap-cloud-for-local pattern in real codebases.
Your first local LLM, end to end
Install Ollama, pull Llama 3.2 3B, chat, hit the OpenAI-compatible API, and troubleshoot the five things that go wrong on first install. By the end of this post you have a working local LLM.
Every machine can run a local LLM (here's what fits)
Per-tier guide: 8GB integrated graphics, 16GB MacBook Air, 8/12/16/24/32GB VRAM PCs, 24/32/64/128/192GB Macs. Specific models, specific tok/s, specific configs. Every tier runs something useful.
System requirements by OS for local LLMs
What macOS, Linux, and Windows each need to run a local LLM in 2026. Native Windows now works smoothly; WSL2 for Linux power users; Mac is the smoothest path; Linux gives you the most knobs.
Picking a local model by task
The 2026 open leaders by task: coding (Qwen 2.5 Coder, DeepSeek-Coder), chat (Llama, Qwen, Mistral), small-model renaissance (Phi-3, Gemma 2), structured output, multimodal, embeddings.
Streaming, throughput, and the KV cache
TTFT vs tok/s, why streaming feels faster, and the KV cache that makes the 1000th token cost the same as the first. KV cache quantization (Q8/Q4 KV) and why it should be your default.
Quantization, distillation, pruning: making models fit
Three ways to shrink an LLM. Quantization (Q2-Q8 with K-quants in GGUF), distillation (teacher to student), pruning. Why Q4_K_M is the community default and what each lever costs.
The local-LLM vocabulary
Parameters, B, dense vs MoE, base vs instruct, tokens, context window, chat template, GGUF, quantization suffixes. After this post you can read any HuggingFace model card.