What local AI model can your PC run?
Models run inside a free app — Ollama or LM Studio. Install either first, then grab a pick below.
No GPU · 8 GB RAM
Typical office laptop. Small models only — genuinely useful for questions, drafting and one-command-at-a-time help.
10–25 words/sec on a typical laptop CPU — roughly reading speed.
The best small model of 2026: strong instruction-following and tool calling for its size.
ollama run qwen3.5:4bSnappier than the 4B on weak or older CPUs.
Noticeably shallower answers — the escape hatch if the 4B feels sluggish.
ollama run qwen3.5:2bDon’t — on 8 GB it swaps to disk and crawls below reading speed.
The upgrade everyone tries first. Wait until you have 16 GB of RAM.
ollama run qwen3.5:9bNo GPU · 16 GB RAM
Most modern laptops without a gaming GPU. The 4B flies; a 9B becomes possible if you’re patient.
Reading speed or better, with enough RAM headroom to keep your browser open.
Start here. 256k context, handles follow-up questions well.
ollama run qwen3.5:4bSimilar speed to the Qwen 4B; more natural conversational tone, accepts image input.
Grab the ~3.5 GB Q4 build in LM Studio — the default Ollama pull of this model is 9.6 GB.
Gemma 4 E4B (Q4)8–15 words/sec — a word at a time, but noticeably smarter answers.
Use when quality matters more than speed.
ollama run qwen3.5:9bNo GPU · 32 GB+ RAM
Desktop-class RAM without a real GPU. Mixture-of-experts models change the math here — big brains, small per-word cost.
MoE design: only ~3B of the 35B are active per word, so CPU-only stays usable.
MoE on CPU is the experiment worth trying at 32 GB.
Qwen3.6 35B A3B (Q4)Comfortable at this RAM level — the reliable pick.
The safe choice while the MoE experiments shake out.
ollama run qwen3.5:9b4–6 GB GPU
Entry gaming laptops and older cards (GTX 1650/1060, RTX 3050). Small models feel instant here.
25–40 words/sec fully on the GPU — feels instant.
Fits entirely in VRAM with room for context.
ollama run qwen3.5:4bFast and personable; image input included.
Q4 build in LM Studio, not the 9.6 GB default Ollama pull.
Gemma 4 E4B (Q4)8 GB GPU
The most common gaming card class (RTX 3060 Ti/4060, RX 6600). A 9B model fits fully and flies.
Fast — fully on GPU with context to spare.
The sweet-spot model for this card class.
ollama run qwen3.5:9bInstant-feeling; handles images too.
Runs alongside other GPU apps comfortably at this size.
Gemma 4 E4B (Q4)10–12 GB GPU
RTX 3060 12GB / 4070 class — a small card that punches far above its weight.
33–36 words/sec measured on an RTX 3060 12GB — near-cloud feel, fully offline.
The trick: LM Studio’s “Force Expert Weights onto CPU” splits it across GPU + RAM. Needs 32 GB system RAM.
Qwen3.6 35B A3B (Q4)~12 words/sec under plain Ollama — slower, but setup is one command.
The community-proven pairing for coding agents — one command and you’re running.
ollama run qwen3-coder:30bVery fast — nothing spills to system RAM.
The zero-fuss fallback if the big MoE setups feel like too much.
ollama run qwen3.5:9b16 GB GPU
RTX 4080 / 4070 Ti Super class. 27B-class models come into range.
~34 words/sec measured on a 16 GB card in a real OpenCode field test.
Q3 quantization to fit — quality holds up in practice.
Qwen3.5 27B (Q3)Fast — the same GPU + RAM split trick, with less spilling than on 12 GB cards.
Needs 32 GB system RAM for the offloaded experts.
Qwen3.6 35B A3B (Q4)Fully on GPU, quick and personable.
The everyday pick when you’re not doing agent work.
ollama run gemma4:12b24 GB GPU
RTX 3090/4090 class. The best local coding models run fully on GPU — no compromises.
Fast, fully on GPU, big context headroom.
Widely rated the best local agentic coder of mid-2026.
Qwen3.6 27B (Q4)Entirely on GPU at this VRAM — extremely quick.
MoE speed without the offload tricks.
Qwen3.6 35B A3B (Q4)Clean tool calls and a strong all-rounder; tight fit at 24 GB.
Slightly slower than the Qwen MoE in same-rig tests.
Gemma 4 26B A4B (Q4)32 GB+ GPU / multi-GPU
Workstation territory (RTX 5090, dual cards). Higher-precision builds of the best models.
The best local coder with less quantization loss.
Headroom for maximum context windows and higher precision at 32 GB+.
Qwen3.6 27B (Q6/Q8)Apple Silicon · 16 GB unified
M-series Macs share memory between CPU and GPU — your RAM effectively is your VRAM. Use LM Studio (MLX builds).
Fast on any M-series chip.
MLX builds are optimized for Apple Silicon — prefer them over GGUF.
Qwen3.5 9B (MLX)Instant-feeling.
Leaves plenty of memory for your actual work.
Gemma 4 E4B (MLX)Apple Silicon · 32 GB unified
M-series Pro class. MoE models love unified memory — no GPU/RAM split to manage at all.
Fast — the whole model sits in unified memory.
The standout pick for 32 GB Macs.
Qwen3.6 35B A3B (MLX Q4)Solid speed, proven quality.
The alternative if you prefer steady, predictable memory use.
Qwen3.5 27B (MLX Q4)Apple Silicon · 64 GB unified
M-series Max class. High-precision 27B builds and fast MoE — a serious local AI workstation.
Fast and precise.
Room for huge context windows at this memory level.
Qwen3.6 27B (MLX Q6)Extremely quick.
The speed pick.
Qwen3.6 35B A3B (MLX Q4)Apple Silicon · 128 GB+ unified
M-series Ultra class. Frontier-adjacent models at home.
Close to cloud-model quality, fully offline.
Ultra-class memory bandwidth keeps this big MoE quick.
Qwen3.6 122B A10B (MLX Q4)Changelog
- 2026-07-13 — Initial version, seeded from the July 2026 research pass.
- 2026-07-13 — Review pass: filter now checks your RAM against each pick (with warnings), honest below-minimum guidance, dedicated-vs-shared GPU memory clarified, clearer run commands.
How this list works: curated picks per hardware tier — not every model that exists. “Worth your time” means it runs at a speed and quality that won’t make you regret the download. The bar on each card shows the model’s download size against your tier’s memory. Speeds are from measured community reports on comparable hardware; picks are reviewed on a schedule as new models release — see the changelog. Sizes are the standard (Q4) build unless noted: Q4/Q6/Q8 are compression levels (higher = bigger but slightly sharper). “MoE” models activate only a small part of themselves per word — that’s why some huge ones run on modest hardware. Why mostly Qwen and Gemma right now? They currently lead open-model quality per gigabyte at these sizes — the picks change when the data does.
To run a pick: Ollama (paste the command in a terminal) or LM Studio (a friendly app — search the model name in its built-in browser).