What Can I Run AI?

HomeCan I run it?

Can I run this model on that GPU?

248 curated answers for the models and machines people actually search for. Each page leads with a yes or no, the quant that fits, VRAM needed vs usable, an estimated tokens/sec range, and what to try instead. These are estimates from our formula, not lab benches.

Want every model on one machine? Use the hardware directory or the VRAM calculator. The pages below stay short on purpose — they do not repeat the full 656-model tables.

Qwen3-32B

Alibaba’s dense 32B Qwen3, a common “best quality that still fits a 24 GB card” pick. About 23.8 GB at Q4.

Qwen3-30B-A3B

A Qwen3 Mixture-of-Experts with 30B total weights and about 3B active per token, so it needs 30B-class memory but can decode closer to a small dense model. About 22.3 GB at Q4.

Qwen3.8-27B

Alibaba’s denser Qwen3.8 27B, a newer mid-size chat model in the same memory class as Gemma 3 27B. About 20.4 GB at Q4.

Gemma 3 27B

Google’s larger Gemma 3 instruct model, in the same memory band as other mid-20B dense models. About 20.1 GB at Q4.

Gemma 4 31B

Google’s dense Gemma 4 31B. Close in memory to Qwen3-32B, with a different training mix. About 22.8 GB at Q4.

gpt-oss-20b

OpenAI’s open 20B-class gpt-oss model. Only a few billion parameters are active per token, so decode can be quicker than a dense 20B. About 15.6 GB at Q4.

Mistral Small 24B

Mistral’s 24B Small instruct model, a dense mid-size alternative to the 27B–32B class. About 17.7 GB at Q4.

Llama 3.1 8B

Meta’s widely used 8B instruct model, the default small local chat model for a few years running. About 6.6 GB at Q4.

Qwen3-8B

Alibaba’s 8B Qwen3. A strong small dense model, and a common upgrade from older 7B–8B weights. About 6.7 GB at Q4.

Qwen3-14B

Alibaba’s 14B Qwen3. It sits in the sweet spot for 12–16 GB cards that want more quality than an 8B. About 11.3 GB at Q4.

Gemma 3 12B

Google’s mid-small Gemma 3, sized for 12–16 GB cards that want more than a 4B or 8B. About 9.5 GB at Q4.

Phi-4

Microsoft’s Phi-4, a dense 14B-class model often compared with other 12B–14B instruct weights. About 11.2 GB at Q4.

Llama 3.2 3B

A tiny Meta Llama, aimed at phones and low-VRAM cards. Quality is limited; it is the model you pick when memory is tight. About 3.2 GB at Q4.

Gemma 3 4B

Google’s small Gemma 3 instruct model, a good low-VRAM option with a different flavor from Llama 3.2 3B. About 4.0 GB at Q4.

Llama 3.3 70B

Meta’s flagship 70B instruct model, the usual target for a high-end local chat box. About 50.2 GB at Q4.

Qwen3-Coder 30B

The coding-tuned sibling of Qwen3-30B-A3B. Memory need matches the base 30B-A3B; the weights are trained for code. About 22.3 GB at Q4.

R1 Distill 32B

A 32B DeepSeek-R1 distill on Qwen. Memory need matches Qwen3-32B; the weights are distilled from DeepSeek-R1. About 23.8 GB at Q4.

R1 Distill 70B

A 70B DeepSeek-R1 distill on Llama. Memory need matches Llama 3.3 70B; it is not the full DeepSeek-R1. About 50.2 GB at Q4.

gpt-oss-120b

OpenAI’s open 120B-class gpt-oss model. It needs large unified memory at Q4, with a much smaller active set per token. About 82.5 GB at Q4.

Qwen3-235B

Qwen3’s large Mixture-of-Experts (235B total, about 22B active). It needs a high-memory unified machine at Q4, not a single consumer GPU. About 164.9 GB at Q4.

DeepSeek-R1

The full DeepSeek-R1 teacher model (hundreds of billions of parameters). No single device we track holds it at Q4. About 478.2 GB at Q4.

DeepSeek-V3.2

DeepSeek’s large V3.2 model. Like full R1, it is far beyond a single consumer GPU or current Mac Studio at Q4. About 478.8 GB at Q4.

Mixtral 8x22B

Mixtral 8x22B, a large Mixture-of-Experts. Total weights need a high-memory box; only a fraction is active per token. About 99.0 GB at Q4.

GLM-4.5-Air

Zhipu’s GLM-4.5-Air, a large model in the roughly 80 GB Q4 class — above a 24 GB card, below full DeepSeek-R1. About 78.0 GB at Q4.

R1 Distill 7B

A 7B DeepSeek-R1 distill on Qwen. It is a reasoning-tuned small model, not the full 671B R1. About 6.3 GB at Q4.

R1 Distill 14B

A 14B DeepSeek-R1 distill on Qwen. Same memory class as other 14B dense models, with reasoning-tuned weights. About 11.3 GB at Q4.

Quick answers

What do these pages answer?
Each page answers whether a named open model fits on a named GPU or Mac, at which quant, how much memory it needs versus what the machine can use, an estimated tokens-per-second range, and what to try if it does not fit.
Are the tokens-per-second numbers benchmarks?
No. They are estimates from memory bandwidth and our published efficiency bands, not lab measurements. Mixture-of-Experts figures are an upper bound.
How do you decide if a model fits?
We estimate memory as weights plus about 15% and 1 GB for a short context and runtime. It fits if that is at most 90% of usable memory, tight up to 100%, and no above that.
Why is my GPU or model missing?
This index is curated for popular searches, about 248 pairs, so the pages stay short and distinct. The hardware and model directories still cover every device and model we track.

New models, and what they need to run

One short email a week: fresh open models with their VRAM numbers, plus hardware worth watching.

Or subscribe here