RelicRELIC
AI on your machine

The best local LLM depends on your hardware

There is no single best local LLM, and anyone who names one without asking about your hardware is guessing. The right model is the smallest one that does your task well and fits your memory. This is a buyer's guide organized the way the decision actually works: by hardware tier first, then by task.

Jordan Gibbs July 10, 2026 9 min read

Search for the best local LLMand you get a stream of confident, contradictory answers, most of them out of date within weeks. That is because the question is framed wrong. “Best” is not a property of a model in isolation. It is the intersection of what your machine can hold, what you are trying to do, and how much quality you actually need. Get those three right and the choice narrows to two or three obvious candidates.

So this guide does not crown a winner. It gives you a method. If you have not yet met the vocabulary of parameters and quantization, start with our pillar on Ollama models, which lays out the memory math this piece builds on. Then come back here to turn that math into a pick.

The one rule that beats every leaderboard

Choose the smallest model that clears your bar. A smaller model loads faster, leaves memory for your actual work, generates quicker on modest hardware, and often does the everyday job just as well as a giant you have to fight to run. People burn a lot of time chasing the largest model their RAM will technically tolerate, then wonder why their laptop is unusable. Start small, and only size up when the small one demonstrably fails on your real prompts.

Pick by hardware tier first

Your memory sets the ceiling, so this is where every honest recommendation starts. As of mid-2026, for a 4-bit (Q4) model, budget roughly 0.5 to 0.6 GB per billion parameters plus context overhead. That yields these tiers.

8 GB laptop

You are in the 7 to 8 billion parameter class at Q4, and that class earns its keep. A good 7 to 8B model handles chat, drafting, summarizing, and light coding help perfectly well. Pick a well-supported general family and leave headroom for your browser and editor. Do not try to cram a 13B model in here; it will swap to disk and crawl.

16 GB laptop or desktop

The 13 to 14 billion class opens up, and it is a real step up in coherence on harder questions. You can also run a 7 to 8B model with a much longer context window, which is often the better trade when you work with large documents.

24 GB or more GPU

With a discrete card, what counts is VRAM, not system RAM, because the model runs fastest when it fits entirely on the GPU. At 24 GB of VRAM you can hold larger models, or run mid-size ones at higher precision (Q8) for a quality bump. This is the tier where local starts to feel close to hosted for many tasks.

Apple Silicon

Apple’s unified memory is the quiet advantage here. Because the CPU and GPU share one pool, the whole number on the spec sheet is available to the model. A machine with 32 or 64 GB of unified memory can run models that would need an expensive discrete card elsewhere, which is why Apple Silicon laptops have become a favorite for local inference.

A rough memory-to-model map for Q4, mid-2026: 8 GB sits around 7 to 8B, 16 GB around 13 to 14B, 24 GB+ of VRAM or unified memory reaches the 30B and up class. Treat these as starting points, then confirm with ollama ps once a model is loaded.

Then narrow by task

General chat and writing

Any current general-purpose family in your tier will do. This is the least demanding category, and the differences between good models are small enough that you should just pick one with broad support and move on.

Coding

Use a variant tuned for code, and lean toward the largest one your memory allows, since coding punishes weak models more visibly than chat does. Quality here varies enough between families that testing on your own code is not optional.

Embeddings and RAG

For retrieval and semantic search, use a dedicated embedding model. Embedding models are small and fast, and their job is to convert text into vectors so you can find things by meaning. Pair one with a chat model rather than expecting a single model to do both.

Long context

If you routinely feed in long documents, favor a model with a large context window and remember that context consumes memory on top of the weights. Sometimes the right call is a smaller model with a longer window rather than a bigger model with a short one.

Quantization: Q4 versus Q8, and why GGUF

Quantization is the lever that decides how much model fits your memory. GGUF is the file format the common local runners use, and within it you will see quantization tags. The two you meet most:

  • Q4 (about 4-bit): roughly half the memory of Q8, with a small and usually acceptable quality drop. The default for a reason. Choose this when you want the largest capable model that fits.
  • Q8 (about 8-bit): higher fidelity, roughly double the memory. Worth it when you have headroom and want the last bit of quality out of a given model size, common on GPU and high-memory Apple Silicon setups.

A useful heuristic: a larger model at Q4 usually beats a smaller model at Q8 for the same memory budget. Precision matters, but parameters matter more, so spend your memory on size first and precision second.

How to actually evaluate: test on your own prompts

Public leaderboards are a starting signal you should confirm for yourself. They saturate, they get gamed, and they measure tasks that may have nothing to do with yours. The only benchmark that matters is your work. Keep a small file of the prompts you actually run, ten or fifteen representative ones, and paste them into each candidate model. The winner is obvious within minutes, and it is often not the one the leaderboard favored.

a quick A/B on your own prompts
# pull two candidates
ollama pull llama3.1:8b
ollama pull qwen2.5:7b

# run the same real prompt through each and compare
ollama run llama3.1:8b
ollama run qwen2.5:7b

Judge on the outputs you care about: does it follow instructions, is it accurate on your domain, does it stay coherent over a long reply. A model that tops a chart but ignores your formatting requests is not the best model for you.

Tooling beyond Ollama

Ollama is the easiest on-ramp, but it is not the only one, and knowing the neighbors helps you pick the fit.

  • LM Studio is a graphical app for browsing, downloading and chatting with local models, friendlier if you prefer a window to a terminal.
  • llama.cpp is the underlying inference engine that several of these tools wrap. Using it directly gives the most control over quantization and runtime flags, at the cost of more setup.

All three run the same GGUF models, so the model math in this guide applies whichever you choose. The tool is a matter of taste; the memory ladder is not.

Picking your model in five lines

  • The best model is the one that fits your memory and does your task; that changes per person.
  • Pick your tier by RAM or VRAM, then the smallest model in it that clears your bar.
  • Prefer a larger model at Q4 over a smaller one at Q8 for the same budget.
  • Ignore the leaderboard and test on your own prompts; the answer changes.
  • Ollama, LM Studio and llama.cpp all run the same models, so choose the interface you like.

Once you have your pick, the pillar on Ollama models has the exact commands to pull, run and manage it.

Written by
Jordan GibbsFounder, Relic

Jordan Gibbs is the founder of Relic, an end-to-end encrypted, permanent, searchable memory for everything you copy. He writes widely about AI, agents, and practical tooling on Medium, where he is read by tens of thousands, and builds privacy-first software. Here he covers how everyday tools like the clipboard actually work, and how to use them without handing your data to someone else.

MediumGitHubLinkedIn
Part of
AI on your machine
Keep reading
Pillar·10 min

Ollama models: how to choose and run them

Ollama runs language models on your own hardware with almost one command. The catch is picking one your machine can hold. How to read a model listing, do the RAM math, and run it.

Read
8 min

Claude Code vs Cursor: an even-handed comparison

Claude Code is a terminal agent you delegate tasks to; Cursor is an AI-first editor you stay inside. A fair comparison of form factor, interaction and where each one shines.

Read
8 min

Prompt injection examples, explained for defenders

Prompt injection turns untrusted data into instructions a model obeys. The categories of prompt injection examples, why they work, and how to defend without pretending it is solved.

Read
local aillmcomparison