What is the best local LLM for 8GB of VRAM?

Qwen 3.5 9B at Q4_K_M needs about 6.4 GB with 8K context and still fits a much longer context on 8 GB. Gemma 4 E4B (5.7 GB) is the pick if you want image and audio input.

What is the best local LLM for 24GB of VRAM?

Qwen 3.6 27B (16.9 GB at Q4_K_M) for quality and very long context, or Qwen 3.6 35B-A3B (21.0 GB) for speed. Gemma 4 31B is a strong alternative.

Which local model is closest to ChatGPT?

At home, the closest are the large MoE models: gpt-oss-120b and Qwen 3.5 122B-A10B on 96–128 GB machines, or Qwen 3.5 397B, DeepSeek V4 Flash and Kimi K2.6 on 256 GB+ systems. On a single 24 GB GPU, the 27–35B Qwen 3.6 models are the strongest general choice.

What is the best local LLM for 16GB of VRAM?

gpt-oss-20b in its native format needs about 13.9 GB and is very fast. Qwen 3 14B at Q5_K_M or Qwen 3.5 9B at Q8_0 are strong dense alternatives.

Are local LLMs as good as ChatGPT?

The largest open models are close on many tasks, but they need server-class hardware. Models that fit a single 24 GB GPU are very capable for coding, writing and analysis, though behind the best hosted models on the hardest problems.

Best Local LLMs for Every VRAM Size

Updated October 14, 2026 · Figures calculated live from our model data

Pick the largest good model that fits your memory with the context you need. In 2026 that means Qwen 3.5 9B or Gemma 4 E4B on 8 GB, gpt-oss-20b on 16 GB, Qwen 3.6 27B or Qwen 3.6 35B-A3B on 24 GB, and gpt-oss-120b or Qwen 3.5 122B-A10B on 96–128 GB machines.

Quick picks

All figures are total memory at Q4_K_M (or the official native format) with an 8K context and FP16 KV cache. "Max context" is the longest context that fits the tier at that quantization.

VRAM	Pick	Needs	Max context	Why
8 GB	Qwen 3.5 9B	6.4 GB	44K	Hybrid attention keeps long context cheap
8 GB	Gemma 4 E4B	5.7 GB	124K	Text, image and audio input
12 GB	Qwen 3 14B	10.5 GB	13K	Strong dense 14B; thinking mode
12 GB	Ministral 3 14B	9.7 GB	18K	Vision, Apache 2.0
16 GB	gpt-oss-20b	13.9 GB	55K	Native MXFP4, very fast
16 GB	Qwen 3.5 9B at Q8_0	10.2 GB	163K	Near-lossless 9B with a long context
24 GB	Qwen 3.6 27B	16.9 GB	98K	Dense 27B, very long context
24 GB	Qwen 3.6 35B-A3B	21.0 GB	88K	MoE: small-model speed
24 GB	Gemma 4 31B	19.8 GB	43K	Largest dense Gemma 4
32 GB	Qwen 3.8 27B at Q6_K	22.3 GB	131K	Near-lossless 27B with room for context
48 GB	Llama 3.3 70B	43.2 GB	14K	Classic dense 70B
64 GB	Qwen3-Next 80B-A3B	46.3 GB	256K	Fast 80B MoE
96 GB	gpt-oss-120b	62.0 GB	128K	Native MXFP4, 5.1B active
96 GB	Qwen 3.5 122B-A10B	70.2 GB	256K	122B MoE, tiny cache
96 GB	GLM-4.5-Air	62.1 GB	128K	Popular for coding and tools
256 GB+	Qwen 3.5 397B-A17B	226 GB	256K	Mac Studio or server class

How we chose

We favour recent models with permissive licences and an architecture that makes good use of memory. Benchmarks move every month, so treat these as strong defaults rather than a ranking. For any other model, check the full model list or the VRAM tier pages, which list every model that fits each size.

Why newer models fit better

The big change in 2026 is attention design. Qwen 3.6 27B keeps a KV cache on only 16 of its 64 layers, so 32K of context costs 2.1 GB, against 8.0 GB for Qwen 3 32B. Gemma 4 uses sliding windows on five of every six layers. That is why the 24 GB picks above can hold contexts that were out of reach a year ago.

Dense or mixture-of-experts?

Mixture-of-experts (MoE) models such as Qwen 3.6 35B-A3B read only a few billion parameters per token, so on an RTX 3090 24GB it generates ~138 tokens/s against ~36 for the dense Qwen 3.6 27B. Dense models of similar size are often a little stronger per gigabyte. See MoE models and VRAM for the trade-offs.

Coding, images and video

For coding agents on one GPU, see the best coding models for 24 GB.
For image generation, Z-Image Turbo runs on 8 GB and FLUX.1 [dev] on 12 GB with GGUF weights; see running Flux on low VRAM.
For video, start with Wan 2.2 TI2V 5B or LTX-2; see VRAM for AI video.

Picks by task

General chat and writing: Qwen 3.6 27B on 24 GB, Qwen 3 14B on 12–16 GB, Qwen 3.5 9B on 8 GB.
Coding: Qwen 3.6 35B-A3B or Qwen3-Coder 30B-A3B for speed, Qwen 3.6 27B for quality. See coding models for 24 GB.
Images in, text out: Gemma 4 31B, Gemma 4 E4B and Ministral 3 14B all accept images.
Long documents: hybrid and sliding-window models keep the cache small: Qwen 3.8 27B, Gemma 4 26B-A4B.
Fastest on a Mac: mixture-of-experts models such as gpt-oss-120b and Qwen3-Next 80B-A3B.
Reasoning: thinking modes in Qwen and gpt-oss, or DeepSeek R1 Distill Qwen 32B. Reasoning models write long hidden answers, so budget extra context.

Licences

Most of our picks use permissive licences that allow commercial use: Apache 2.0 (Qwen, Gemma 4, Mistral Small 4, Ministral 3, gpt-oss) and MIT (GLM, DeepSeek, MiniMax). Llama models use Meta's community licence, which has an acceptable-use policy and a clause for very large companies. Check the licence on the model page before shipping a product.

Where to download

Ollama and LM Studio have built-in catalogues; search for the model name and pick the quantization tag. For llama.cpp, GGUF files are on Hugging Face from the model makers and from well-known community quantizers. Prefer uploads that state which llama.cpp version and importance matrix they used, and check the file size against our GGUF size calculator: a file much smaller than expected is usually a lower quantization than its name suggests.

How often this changes

Open models improve every few months. We add new releases to the model list within days and log every addition in the changelog. The memory figures on this page update automatically when the data changes, so the table above always reflects our current numbers.