← All guides

Which Local LLM Fits Each RAM Tier (2026): File Sizes and Ollama Tags

Pick a model by its file size, not by its parameter count. The table below gives the file size of each pick at each RAM tier, from 16GB to 256GB, and an Ollama tag that resolves. Every size comes from the Hugging Face file tree or the Ollama registry, read on 2026-11-21.

Correction, 2026-11-21 — This page used to show a "30-day benchmark" with success rates, tok/s and peak RAM. Some of its peak-RAM numbers were smaller than the model files, so those runs could not have happened as written. We removed every number from that run. This page now lists file sizes only, each with its source.

The rule: the file must be smaller than your RAM

A model loads its whole file into memory. Your OS, your apps and the context window (KV cache) need memory too. So pick a file that is clearly smaller than your RAM. A file that is about the same size as your RAM does not fit.

Parameter count does not tell you the file size. gpt-oss 20B is a 12-14 GB file because it ships in MXFP4. A dense 27B model at Q8 is 29 GB. Read the file size.

Fit table: one or more picks per RAM tier

Sizes are GB of 10^9 bytes. “Registry” means the Ollama registry manifest. “HF” means the Hugging Face file tree. All read on 2026-11-21.

RAM tierModel and quantFile sizeOllama tagSource
16GBQwen 3.5 9B Q8_09.53 GBqwen3.5:9b-q8_0Registry
16GB (tight)gpt-oss 20B MXFP413.79 GBgpt-oss:20bRegistry
24GBQwen 3.6 27B Q4_K_M16.84 GBqwen3.6:27bRegistry
24GBQwen 3.6 35B-A3B Q4_K_M21.72 GBqwen3.6:35bRegistry
32GBQwen3.8 27B UD-Q6_K21.98 GBhf.co/unsloth/Qwen3.8-27B-GGUF:UD-Q6_KHF
32GBGemma 4 26B-A4B17.07 GBgemma4:26b-a4bRegistry
32GBgpt-oss 20B Q8_012.11 GBhf.co/unsloth/gpt-oss-20b-GGUF:Q8_0HF
32GBNemotron Cascade 2 30B24.27 GBnemotron-cascade-2:30bRegistry
48GBQwen 3.6 27B Q8_029.05 GBqwen3.6:27b-q8_0Registry
48GBQwen 3.6 35B-A3B UD-Q6_K29.31 GBhf.co/unsloth/Qwen3.6-35B-A3B-GGUF:UD-Q6_KHF
64GBQwen3.8 27B Q8_029.05 GBhf.co/unsloth/Qwen3.8-27B-GGUF:Q8_0HF
64GBQwen 3.6 35B-A3B Q8_036.90 GBhf.co/unsloth/Qwen3.6-35B-A3B-GGUF:Q8_0HF
96GBgpt-oss 120B MXFP465.37 GBgpt-oss:120bRegistry
96GBQwen 3.5 122B-A10B Q4_K_M81.37 GBqwen3.5:122bRegistry
128GBgpt-oss 120B MXFP465.37 GBgpt-oss:120bRegistry
128GBLlama 4 Scout Q4_K_M67.44 GBllama4:scoutRegistry
128GBQwen 3.5 122B-A10B Q4_K_M81.37 GBqwen3.5:122b-a10b-q4_K_MRegistry
192GBDeepSeek V4 Flash UD-IQ4_XS136.66 GBnone checkedHF (unsloth/DeepSeek-V4-Flash-0731-GGUF)
256GBLlama 4 Maverick UD-Q3_K_XL179.55 GBnone checkedHF (unsloth/Llama-4-Maverick-17B-128E-Instruct-GGUF)
256GBLlama 4 Maverick IQ4_XS214.32 GBnone checkedHF (same repo)

qwen3.5:122b and qwen3.5:122b-a10b-q4_K_M resolve to the same 81.37 GB file. The 192GB and 256GB files are split GGUFs. We list no pull tag for them because we did not check one.

Models that do not fit the tier people expect

ModelFile sizeCommon claimFact
gpt-oss 120B65.37 GB”Fits 64GB”The file alone is larger than 64GB. Use 96GB or more.
Llama 4 Scout Q4_K_M65.36-67.44 GB”Fits 64GB”Too large for 64GB. The IQ4_XS file is 57.65 GB, with little room for context.
Llama 4 Maverick Q4_K_M242.77 GB”Fits 128GB at Q4”The file is far larger than 128GB. Even UD-Q3_K_XL is 179.55 GB.
DeepSeek V4 Flash128.21 GB at UD-Q3_K_XL, 136.66 GB at UD-IQ4_XS”Fits 128GB at Q4”The Q4 file is larger than 128GB.

What this page does not measure

This page gives no speed, quality or tool-call success numbers. Speed depends on memory bandwidth, quant and context length. For those, read the tier guide for your RAM. Each guide cites its own sources.

How to check a file size yourself

For an Ollama library tag, open https://ollama.com/library/<model>/tags. Each tag shows its size. For an hf.co tag, open the repo’s “Files” tab on Hugging Face and read the size of the .gguf file. Split files list each part, so add the parts.

A laptop for the 32GB tier picks

36GB of unified memory holds every 32GB-tier file in the table above.

Amazon affiliate links — we earn a small commission at no cost to you.

Read next

Best Local LLM for Mac Studio M2 Ultra (2026)
Best local LLM for the Mac Studio M2 Ultra. 2026 picks for 64GB, 128GB, 192GB variants. gpt-oss 120B, Mistral Small 4 (119B-A6B), Llama 3.3 70B Q8, and quad-model OpenClaw setups.
Can I Run Llama 3.3 70B With 64GB RAM?
Llama 3.3 70B can fit at Q4 with about 40GB memory, but 64GB RAM is tight for long OpenClaw runs. Use this compatibility guide.
32GB vs 64GB RAM for Local LLMs: Which Should You Buy?
Should you buy 32GB or 64GB RAM for local LLMs and OpenClaw? A direct comparison with model tiers, tradeoffs, calculator links, and upgrade rules.
Best Local LLM for 32GB RAM (2026): Chat and Coding Picks
Best local LLM for 32GB RAM: Qwen3.8 27B at UD-Q6_K (20.47 GiB). gpt-oss 20B is 11.28 GiB, not 22GB. What fits under the ~24 GiB Mac GPU limit, Ollama commands, what to skip.