"Local LLMs 2026 — Ollama vs LM Studio, and What Actually Runs on Your Machine"

Install to model selection, with picks for 8 GB / 16 GB / 24 GB / 64 GB systems on Mac and Windows

ํ•ต์‹ฌ ์š”์•ฝ

  • Audience: Anyone who wants to run LLMs on their own computer for cost, privacy, or learning — both general users and developers.
  • What you'll get: 1) Ollama vs LM Studio — which to pick, 2) what runs on 8/16/24/64 GB systems, 3) Mac and Windows install steps, 4) realistic use cases and limits, 5) honest comparison vs cloud.
  • One-liner: First-time GUI → LM Studio. Automation/server/dev → Ollama. Both free. Apple M-series surprises at 8 GB and shines at 64 GB.

1. Why local

Factor Cloud Local
Cost Per-token Free after install (electricity, hardware)
Privacy Sent to vendor Never leaves your machine
Offline Needs network Works
Quality GPT-5.5 / Claude 4.7 / Gemini 3 Llama 3 / Qwen 3 / Mistral (open-source SOTA)
Speed Depends on model + network Depends on your hardware (~30–40 tok/s on a typical laptop)
Multimodal Rich Text + some vision
Context 1M+ 8K–128K

Local fits when: - Sensitive data (medical, legal, contracts). - Heavy automation that's making cloud costs sting. - Offline environments. - Learning / experimentation (model comparison, fine-tuning).

Cloud still wins when: - You need top-end quality on hard tasks. - Multimodal (image/audio in/out). - Casual usage.

2. Ollama vs LM Studio

2.1 Core differences

Ollama LM Studio
Interface CLI (terminal) GUI desktop app
Audience Devs / automation Non-devs / experimenters
Model library Curated (ollama.com/library) HuggingFace integration
Apple Silicon Metal (good) MLX (1.6× faster on M3 Ultra)
GPU auto-detect NVIDIA / AMD / Apple Manual config
Docker Official image ❌
API server Built in (localhost:11434) Toggle in app
n8n / LangChain First-class Works via OpenAI-compatible API
Free ✅ ✅

2.2 Which to pick

  • GUI exploration / model browsing → LM Studio
  • CLI / automation / server → Ollama
  • Both is fine — they don't conflict. Many start with LM Studio, then add Ollama for dev work.

2.3 Speed (M3 Ultra, Gemma 3 1B)

  • LM Studio (MLX): 237 tok/s
  • Ollama (Metal): 149 tok/s

Ollama is adopting MLX, but as of April 2026 LM Studio is faster on Apple Silicon.

3. What runs on your machine

3.1 Model picks by RAM/VRAM (Q4_K_M)

System Recommended model Memory used Use
8 GB VRAM (or 8 GB Apple unified) Qwen 3.5 9B (default rec) 6.6 GB General chat, summary, translate
8 GB Llama 3.1 8B 5–6 GB English-first chat
8 GB Mistral 7B 4.5 GB Lighter tasks
16 GB Qwen 3 14B Q5_K_M 10.2 GB High-quality chat
16 GB Mistral Small 24B Q4 13.4 GB Stronger reasoning
24 GB Mixtral 8x7B / 35B Q4 18–22 GB Production-tier work
48+ GB Llama 3.3 70B Q4 / Qwen 2.5 72B Q4 40–44 GB Near-cloud quality
Apple M-Mac 64 GB+ Llama 3.3 70B / Qwen 72B uses unified memory M5 Max 64 GB ≈ H100-class

Sources: LocalLLM.in VRAM guide, GitHub ollama/ollama.

3.2 Quantization in one line

Q4_K_M = 4-bit compression. ~75% smaller than FP16, with ~5–10% quality loss. The standard default. - Q8 / FP16 = full quality (4× memory) - Q5_K_M = slightly better, +20% memory - Q3 = even smaller, but visibly worse

Start with Q4_K_M everywhere; only step up if quality bothers you.

3.3 The Apple Silicon trick

Unified memory lets the GPU use system RAM. A 64 GB MacBook Pro is effectively a 64 GB VRAM machine. 70B models that won't fit on an RTX 4090 (24 GB) run on M3/M4/M5 Max 64 GB.

4. Install — Mac & Windows

4.1 Ollama (CLI)

Mac

brew install ollama

ollama run llama3.1:8b

Windows (10/11, native ARM64 in 2026) 1. Get the installer from ollama.com/download 2. After install, in PowerShell:

ollama run qwen3:9b

Common commands

ollama list           # installed models
ollama pull mistral   # download without running
ollama rm llama3.1    # remove
ollama serve          # API server (default port 11434)

4.2 LM Studio (GUI)

Mac & Windows 1. Download from lmstudio.ai 2. Open → search icon → search a model (e.g., "qwen 3.5 9b") 3. Download → use the "Chat" tab right away 4. (Optional) "Local Server" tab → Start → enables an OpenAI-compatible API

LM Studio shows "Will fit in your RAM/VRAM" on every model — friendlier on hardware questions.

5. Five practical use cases

5.1 Sensitive document analysis

Summarize legal contracts, medical records, internal memos without sending them anywhere. Even Qwen 3.5 9B on an 8 GB Mac is enough.

5.2 Coding assist

Llama 3.1 8B or Qwen 2.5 Coder 7B paired with the Continue VS Code extension — a viable Copilot alternative.

5.3 RAG over your own files

Ollama API + LangChain/LlamaIndex over your PDFs/notes — your own NotebookLM that never phones home.

5.4 Automation (n8n + Ollama)

Swap the Claude Haiku call from Part 7's email-triage workflow with a local Llama → token cost goes to zero.

5.5 Learning / experimentation

Compare model behavior, choose a fine-tuning base, measure quantization impact.

6. Honest limits

  • Quality gap: a 70B local model is close, not equal, to GPT-5.5 / Claude Opus 4.7. Most visible on reasoning, code, and non-English tasks.
  • Multimodal is limited: image/audio in/out is rougher than cloud.
  • Context windows: typically 8K–32K (some 128K). Cloud is at 1M+ now.
  • Heat / fan / power: 70B on a laptop = fans on full. Desktops cost a bit more in electricity.
  • First download is slow: 4–80 GB per model, 30 minutes to 2 hours over Wi-Fi.
  • Weaker safety alignment: open-source models have lighter safety filters — you carry the responsibility.

7. First-week starter plan

Day Task
1 Install LM Studio → download Qwen 3.5 9B → spend an hour chatting; compare with cloud feel.
2 Install Ollama → pull the same model → CLI run → start the API server.
3 Connect your Python or n8n code via the OpenAI-compatible API. Run it without any cloud key.
4–5 Try RAG on one PDF (LM Studio built-in or LangChain).
6–7 Evaluate. Tally savings vs lost capabilities → keep / mix / drop.

8. Cloud vs local — the honest call

Cloud (Claude / GPT / Gemini) Local (Llama / Qwen / Mistral)
Quality Top-tier Strong (15–30% gap)
Cost Usage-based $0 ongoing, hardware + power
Speed Network-bound Hardware-bound
Privacy Policy-dependent Total
Multimodal Rich Limited
Context 1M+ 8K–128K
Scaling Unlimited Hardware ceiling
Liability Vendor-shared All yours

Practical recommendation: Cloud as your main stack, local for sensitive work and automation. The "all of one" stance loses on both fronts; mixing is the rational play.

Developer notes

  1. Ollama API is OpenAI-compatible: set OPENAI_API_BASE=http://localhost:11434/v1 and reuse the OpenAI SDK as-is. Switch between local and cloud with one env var.
  2. Mind the licenses: Llama 3 (Meta), Qwen (Alibaba), Mistral — each license differs. Some restrict commercial use over MAU thresholds.
  3. n8n integration: use the Ollama Chat node or the OpenAI Chat node with a custom Base URL.
  4. Fine-tuning: 8B models LoRA-tune on 16 GB VRAM. Tools like unsloth and axolotl make this practical.
  5. Benchmarking: lm-eval-harness, MTEB — automate model comparisons on your own dataset.
  6. Production serving: vLLM (concurrent requests), TGI (HuggingFace), llama.cpp server beat Ollama on throughput. Ollama is best for personal use and MVPs.
  7. Edge deployment: Raspberry Pi 5 + 8 GB handles 3B models. Mobile gets Llama 3.2 1B/3B at quantization.

References

This is part 8 of 11 in the AI Basics series. Series complete. Browse the full series in the blog TOC page.

๋Œ“๊ธ€

๋Œ“๊ธ€ ์“ฐ๊ธฐ

์ด ๋ธ”๋กœ๊ทธ์˜ ์ธ๊ธฐ ๊ฒŒ์‹œ๋ฌผ

Agent Memory Engine (2/10) — Building an AI Agent Memory System with SQLite Alone

← 1/10 SQLite Memory Engine f… ๐Ÿ“š Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere ํ•ต์‹ฌ ์š”์•ฝ This post covers how to design an AI agent memory engine (memcore) on a single-file SQLite backend. - Structure: 25 modules, ~6,464 lines, 22 DB tables - Principles: no daemon, single-file portability, no required external dependencies, host-agnostic - Core techniques: 4-layer hybrid search (topic · vector · FTS5 · LIKE), U-tag dialectic-based confidence evolution, ingestion-layer bias rejection rules What This Post Covers One approach to building AI agent memory without a server or cloud vector DB. Specifically: (1) the limitations of text-file-based memory, (2) vector DB alternatives compared and selection criteria, (3) how to layer search tiers, (4) how to evolve confidence scores over time, and (5) how to block biased memory accumulation at the ingestion stage. Design Principles: Why Single-File SQLite Text-file-based ...
Read more »

"ML Foundations (9/9) — PyTorch vs TensorFlow, and the Road to Local LLMs"

์ด๋ฏธ์ง€
← 8/9 Deep Learning Architec… ๐Ÿ“š Series Index (series end) The final part. So far we've focused on models. Now we focus on the tools that actually run them . We'll lay out the philosophical difference between PyTorch and TensorFlow, and trace how Hugging Face transformers , llama.cpp , MLX , and Ollama built the bridge to running large language models on your own machine. By the end you should have the full mental model of "download a pretrained LLM and serve it locally." 0. Learning Objectives Explain the eager-vs-graph difference between PyTorch and TensorFlow. Explain, in graph terms, how autograd automates the backward pass. Use Hugging Face transformers ' AutoModel , AutoTokenizer , and pipeline abstractions. Describe llama.cpp's GGUF quantization, INT4 inference, and CPU-first flow. Describe Apple MLX's use of unified memory and how it differs from PyTorch. Run a local LLM with Ollama and call it through an OpenAI-compatible API. ...
Read more »

"RAG Core Study (14/26) — Evaluation Sets with RAGAS & DeepEval"

์ด๋ฏธ์ง€
Series overview: Series index ๐Ÿ“š Series Index Without an evaluation set, every RAG improvement is a story, not evidence. RAG systems fail in more than one place: retrieval, context selection, answer generation, and grounding. That means "looks better" is not a metric. Part 14 explains how to build a golden dataset , how RAGAS and DeepEval fit into the workflow, and why teams must separate retrieval quality from answer quality if they want tuning decisions to hold up. 0. Prerequisites Part 13 reranking — system quality now depends on multi-stage ranking. Part 1 grounding — a correct-looking answer is not enough. Part 12 Hybrid — multiple retrieval settings need comparison on fixed data. 1. Learning Objectives Build a small but useful golden dataset for RAG. Distinguish retrieval metrics from answer metrics . Use RAGAS and DeepEval in the right roles. Avoid the common traps in judge-model-based evaluation. 2. ํ•ต์‹ฌ ์š”์•ฝ An evaluation set for RAG is ...
Read more »

"ML Foundations (8/9) — Deep Learning Architectures: CNN, RNN, Attention"

์ด๋ฏธ์ง€
← 7/9 Deep Learning Training ๐Ÿ“š Series Index 9/9 PyTorch vs TensorFlow → What did we keep adding on top of an MLP? Vision went CNN, sequences went RNN/LSTM, and eventually both converged on Attention and the Transformer. Each architecture is easier to remember if you read it as what it gave up to gain something else . This part is the one-line summary of the decisive inflection points: LeNet → AlexNet → ResNet → LSTM → Attention → Transformer. 0. Learning Objectives Explain why CNN's convolution, pooling, and weight sharing match images so well. Trace what got unblocked from LeNet → AlexNet → VGG → ResNet as depth grew. State BPTT for RNNs and the vanishing/exploding gradient problem. Write the LSTM gate equations (forget, input, output) with intuition. Write the attention formula in query/key/value form and the scaled dot-product variant. State the precise reasons Transformers replaced RNNs (parallelism + long-range dependencies). 1. ํ•ต์‹ฌ ์š”์•ฝ CNN : weight sharin...
Read more »

"ML Foundations (7/9) — Deep Learning Training: Optimizers, Regularization, Initialization"

← 6/9 Neural Networks ๐Ÿ“š Series Index 8/9 Deep Learning Architec… → Building an MLP, as we did in Part 6, is not the same as getting it to train well . The right optimizer, regularization, initialization, and learning rate — when these line up, deep networks converge. When they don't, the network refuses to learn at all. This part is the catalog of those crafts, with formulas, paper citations, and code in one place. 0. Learning Objectives Compare and write the update rules for SGD, Momentum, Nesterov, and Adam. Explain how Dropout, BatchNorm, and LayerNorm work and where they belong in a model. Derive the variance formulas for Xavier and He initialization and match them to activations. Implement step, cosine, and warmup learning-rate schedules in PyTorch. Explain why gradient clipping is effectively required for RNNs and Transformers. Diagnose the most common training failures (NaN, plateau, overfitting) and apply first-line fixes. 1. ํ•ต์‹ฌ ์š”์•ฝ SGD : \(w \leftarro...
Read more »

OpenClaw to Hermes Migration (2/13) — What to Preserve, Partially Port, or Discard

← 1/13 Current Structure Inve… ๐Ÿ“š Series Index 3/13 What Moves to Hermes → A code review record classifying operational scripts into three tiers: preserve, partial-preserve, and discard. What This Post Covers An asset classification framework (preserve / partial-preserve / discard) for migrating a legacy agent system to a successor A concrete case applying the Strangler Fig + Subset Migration strategy to a memory pipeline Classification results evaluated across LOC, external dependencies, and replaceability axes A method for slimming down a bloated single file (1,310 LOC) using the migration as a natural entry point Problem Definition: Port Everything or Cut Deep? The OpenClaw inventory comprises multiple agents, multiple scheduler/daemon processes, and multiple operational scripts. When moving to the successor system, Hermes, the choice is not binary. A full port carries accumulated technical debt along with it; a minimal port discards hard-won domain logic. Code-revie...
Read more »

AI Agents I Built (5/7) — Building an Automated Blogger API Publishing System

← 4/7 My Life Organizer Agent ๐Ÿ“š Series Index 6/7 Dual → Designing a Markdown-Based Auto-Publish Pipeline with Blogger API v3 + OAuth 2.0 Key Summary Blogger API v3 + OAuth 2.0 enables a pipeline that converts Markdown files into publishable HTML posts blogger_publish.py handles the full flow — frontmatter parsing → body cleanup → HTML conversion → CSS injection → API post — in a single script For bulk publishing, quota management (delay, retry, TOC suppression) is the critical design concern Platform Selection Criteria API access is mandatory for AI agent-driven content publishing. A comparison of platforms: Platform API Cost AdSense Decision WordPress.com REST API Paid plan required Paid only Cost overhead Ghost REST API Self-hosted or paid Manual setup Ops overhead Blogger v3 API Free Native integration Adopted Blogger is free, exposes a REST API, and integrates natively with AdSense. For personal blog automation, this combination i...
Read more »