What LLM Can I Run?

See which LLMs your computer can run. One-click GPU detection checks Llama, Gemma, Qwen, DeepSeek and 50+ models against your VRAM, with speed estimates.

Advertisement

What LLM can I run? Check your hardware against 55 open models at once

This tool starts from the machine you already own and works backwards to a shortlist. You tell it what GPU you have and how much memory it and the system hold; it walks all 55 curated open-weight models — Llama, Qwen3, Gemma, DeepSeek, Mistral, GPT-OSS, GLM, Kimi, Phi 4, Nemotron — and sorts every one into a verdict: runs great, runs with trade-offs, technically runs, or won't fit. It is the inverse of our LLM VRAM calculator, which starts from one model and tells you how much memory that model needs. Use this one when the question is "I have 16 GB, what are my options", and that one when you have already picked the model.

What "Detect my hardware" can and cannot see

The detect button reads two things from your browser: the WebGL unmasked renderer string and, where available, the WebGPU adapter info. Those give a GPU name — strings like ANGLE Metal Renderer: Apple M3 Max or one containing RTX 4070 Ti Super — which is matched against the built-in hardware table to recover bandwidth and typical memory.

Browsers deliberately do not expose memory sizes. navigator.deviceMemory is bucketed and capped at 8 GB, so it is useful only as a floor, and VRAM is not exposed at all. That is why the tool always asks you to confirm the numbers. On Apple Silicon and other unified-memory hardware it makes one extra inference: the WebGPU adapter's maxBufferSize limit is snapped to the nearest shipping memory configuration to pre-select a sensible default. Treat that as a smart guess, not a reading — change it if it is wrong. Everything runs in your browser; no hardware fingerprint is sent anywhere by this screen.

If detection fails, or if you are planning a purchase rather than testing what you own, use "Pick manually" and choose any card from the list — NVIDIA GeForce and datacentre parts, AMD Radeon and Strix Halo, Intel Arc, every Apple Silicon tier, plus plain CPU + DDR4/DDR5 entries.

The quantisation ladder it walks for every model

A model is not one size. The same weights ship in a range of GGUF quantisation formats, and the tool tries them from best quality downwards until one fits your memory. The bytes-per-parameter figures below are averages that include the embedding and output layers, so they match real GGUF file sizes rather than the naive bit-width arithmetic.

FormatBytes per parameterQuality
Q8_01.06Practically indistinguishable from FP16
Q6_K0.82Near-lossless
Q5_K_M0.71Very good, noticeably smaller
Q4_K_M0.61The popular size/quality balance
Q3_K_M0.50Visible quality loss
IQ2_M0.36Significant loss; last resort

Fit is checked at a realistic everyday setting of 8,192 tokens of context, and the memory it needs is weights plus KV cache plus 0.75 GiB reserved for the CUDA or Metal runtime context and buffers. The KV cache is computed from each model's real architecture — layer count, KV head count, head dimension, and whether the layers use sliding-window attention or DeepSeek-style compressed latent attention — not from a rule of thumb.

The five verdicts

VerdictWhat it means
Runs greatFits entirely in GPU or unified memory at Q4_K_M or better, and generates at 15 tokens/sec or more
Runs with trade-offsFits, but only at a reduced quantisation, or fits well and still generates slowly
Technically runsHeavy CPU offload or 2-bit weights — expect single-digit tokens per second
Too slow for youFits, but below the minimum speed you set on the slider
Won't runDoes not fit in VRAM plus system RAM at any quantisation

Worked example: a 16 GB discrete GPU

Point it at a 16 GB card with roughly 717 GB/s of bandwidth (an RTX 4080 is the reference used here) and the shortlist looks like this at 8K context. Speeds are the tool's single-stream estimate at 65% of theoretical bandwidth.

ModelBest fitting quantMemory neededEstimated generation
Llama 3.2 3BQ8_04.8 GiB~107 tok/s
Llama 3.1 8BQ8_09.7 GiB~49 tok/s
Gemma 3 12BQ8_013.6 GiB~34 tok/s
Qwen3 14BQ6_K13.3 GiB~35 tok/s
GPT-OSS 20BQ5_K_M15.2 GiB~169 tok/s
Mistral Small 3.2 24BQ4_K_M15.6 GiB~29 tok/s
Qwen3 30B-A3BQ3_K_M15.7 GiB~190 tok/s
Llama 3.3 70Bdoes not fit in VRAMoffload only

Two things in that table are worth pausing on. First, an 8B model at Q8_0 is a better use of 16 GB than a 14B model squeezed to Q6_K if you care about output quality — you are not obliged to take the biggest model that fits. Second, the mixture-of-experts entries invert the usual ordering. GPT-OSS 20B holds 21.5 billion parameters but reads only about 3.6 billion of them per token, and Qwen3 30B-A3B reads about 3.3 billion of its 30.5 billion. They are large to store and cheap to generate from, so they land near the top of the speed column despite being near the bottom of the fit column. If your bottleneck is capacity, MoE is expensive; if it is speed, MoE is the bargain of the list.

How the answer changes with memory

  • 12 GB (RTX 3060-class): Llama 3.1 8B still runs at Q8_0 in 9.7 GiB. Qwen3 14B drops to Q5_K_M. Mistral Small 24B only fits at IQ2_M, which is the point at which you should prefer a smaller model at a better quant.
  • 24 GB (RTX 3090/4090-class): Qwen3 32B arrives at Q4_K_M in 21.4 GiB, Qwen3 14B moves up to Q8_0, and 70B still needs offload.
  • 32 GB (RTX 5090-class): the same models but faster — 1,792 GB/s roughly doubles a 4090's generation rate on identical weights.
  • 36 GB unified (M3 Max-class): capacity opens up — Llama 3.3 70B squeezes in at IQ2_M — but at ~400 GB/s the big dense models land near 9–10 tok/s. Fits and feels usable are different questions.

Unified memory is not VRAM, and the tool treats it differently

On Apple Silicon, Strix Halo, and integrated graphics, the model shares one pool with the operating system, so the tool budgets 75% of total memory rather than all of it: a 36 GB Mac is assessed as 27 GiB of usable model space. Unified systems also get no CPU-offload fallback, because there is nowhere to offload to — the memory is already the system memory. A discrete GPU, by contrast, can spill layers into system RAM, so those machines are additionally checked against VRAM plus system RAM minus 4 GB for the OS and the runtime.

Why offloaded models collapse in speed

When part of a model lives in system RAM, every generated token has to read those bytes over the memory bus instead of the GPU's. The estimate models that split explicitly: GPU-resident bytes at your card's bandwidth and 65% efficiency, CPU-resident bytes against an assumed 80 GB/s of dual-channel DDR5 at 35% efficiency. That second number is much lower for a reason — dequantisation overhead, unsaturated memory channels and laptop thermal limits mean CPU inference does not achieve anything like the bandwidth the DIMM spec suggests. Mixture-of-experts models take a further penalty on CPU, because switching experts every token defeats CPU caches in a way it does not on a GPU. In practice the RAM-resident share dominates the total time, which is why a model that is "90% on the GPU" is nowhere near 90% of the GPU-only speed.

The minimum-speed slider

Fitting is not the same as being tolerable. The slider (0–20 tokens/sec) reclassifies anything below your threshold into a separate "too slow for you" group, so it stops competing for attention with the models you would actually use. The reference points on the scale: about 5 tok/s is roughly reading speed, 10 feels like smooth chat, 20 feels snappy. Set it to 10 and a long list of technically-runnable 70B configurations quietly disappears, which is usually the honest outcome.

Getting from the shortlist to a running model

Each result carries its chosen quantisation, the memory it needs, the estimated speed, and the largest standard context that still fits alongside the weights on your GPU at that quant — useful when you plan to paste in long documents, since context is the variable that quietly eats the headroom. Every row links out to the VRAM breakdown and the speed detail for that exact model and quant, to an Ollama command builder, and offers a one-click copy of the matching LM Studio command. Results also filter by use case (chat, coding, reasoning, vision) and the whole configuration is captured in the URL, so a shareable link reproduces the same report on someone else's screen.

Where the estimates end

  • Speeds are single-user, batch-size-one estimates from memory bandwidth. Batched serving with vLLM behaves quite differently.
  • The 55-model list is curated. A brand-new fine-tune or an exotic architecture will not be in it.
  • Real throughput varies with inference engine, driver version, and thermals; a laptop that throttles will not hold the estimate.
  • Detection identifies a GPU by name. If your card is not in the matching table, pick the closest entry manually rather than assuming the tool guessed right.
  • Fit is evaluated at 8K context. Long-context work needs more memory than the table shows.

How We Decide What "Runs" on Your Hardware

Whether a model runs is a memory question first and a speed question second.

Memory gate: a model needs room for its weights (parameter count x bytes per parameter at the chosen quantization), its KV cache (grows with context length), and runtime overhead. We check this against your GPU memory first, then against GPU + system RAM combined (CPU offloading).

Speed estimate: for models that fit, generation speed is approximately memory bandwidth divided by bytes read per token, derated to 65% for real-world engine efficiency. Models that need CPU offloading are penalized by system RAM bandwidth, which is the actual bottleneck.

Quantization ladder: we test each model from Q8_0 (near-lossless) down to IQ2_M (heavily compressed), and report the best quality level that fits. A model that only fits at 2-bit quantization is ranked lower because the quality loss is significant.

Reading Your Results: What to Actually Download

Once you know what runs, here is how to act on it:

For chat and general use: pick the largest model in your "runs great" tier. Larger models are smarter; there is rarely a reason to run a 4B model if a 12B model runs well on your hardware.

For coding: coding-tuned models (Qwen3 Coder, Devstral) outperform general models of the same size at code tasks. A 30B MoE coder that runs great beats a 70B general model that runs slowly — latency matters when you are iterating.

For long documents: check the "max context" figure on each card. A model that runs great at 8K context may not handle a 100K-token document. Models with efficient attention (Gemma sliding-window, DeepSeek MLA) degrade least at long context.

Then download it: use the Ollama Command Builder linked on each model card to get the exact pull command for the quantization that fits your hardware.

Frequently Asked Questions

How does the GPU detection work?+

Your browser exposes the GPU model name through WebGL and WebGPU APIs (the same way games detect graphics settings). We match that name against our hardware database to get its specs. Everything happens locally in your browser — nothing is sent to a server. Browsers never reveal memory amounts for privacy reasons, which is why we ask you to confirm your VRAM and RAM.

What do the tiers mean?+

Runs great: the model fits entirely in your GPU memory at good quality (Q4_K_M or better) and generates at comfortable speed. Runs with trade-offs: it works, but needs aggressive quantization (quality loss) or runs slowly. Technically runs: requires CPU offloading or extreme compression — expect single-digit tokens per second. Will not run: the model does not fit in your combined GPU memory and system RAM at any quantization.

What is CPU offloading and why is it slow?+

When a model does not fit entirely in GPU memory, inference engines like llama.cpp can keep some layers in system RAM and run them on the CPU. It works, but system RAM bandwidth (50-90 GB/s) is 10-40x slower than GPU memory bandwidth, so every offloaded layer drags down generation speed. A model that is 40% offloaded typically runs 3-5x slower than one that fits entirely on the GPU.

Why do some huge models run faster than smaller ones?+

Mixture-of-Experts (MoE) models only read a fraction of their parameters for each token. GPT-OSS 120B reads just 5.1B parameters per token, so it generates as fast as a 5B model — but it still needs memory for all 120B parameters. That is why you will sometimes see a 30B MoE model ranked "runs great" while a 32B dense model is marked slow: same memory needs, very different speeds.

What hardware do I need to run a 70B model well?+

For Llama 3.3 70B at Q4_K_M (about 43 GB), you need roughly 48 GB of GPU memory: two 24 GB GPUs (2x RTX 3090 or 4090), one 48 GB workstation card (RTX 6000 Ada), a Mac with 64 GB+ unified memory, or a datacenter GPU. On a 64 GB Mac it runs at about 5-7 tokens/sec; on 2x RTX 4090 with tensor parallelism, around 20-25 tokens/sec.

My GPU was not detected. Why?+

Safari and some privacy-focused browsers report a generic "Apple GPU" or hide the renderer string entirely. Some Linux setups report the Mesa driver name instead of the GPU model. If detection fails, just pick your hardware manually from the vendor list — the results are identical.

Does this account for context length?+

Yes — rankings are computed at 8K tokens of context, a realistic everyday setting. Each model card also shows the maximum context that fits on your hardware. Long contexts need significantly more memory (the KV cache grows linearly), so a model that fits at 8K may not fit at 64K.

What about quantized models I download from Hugging Face?+

The quantization levels we test (Q8_0 down to IQ2_M) correspond to the standard GGUF files you will find on Hugging Face and in Ollama. When a model is ranked "runs great at Q4_K_M", that is exactly the file variant to download. Use our Ollama Command Builder to get the right pull command.

Related tools

This tool is provided for informational and educational purposes only. All processing happens in your browser — no data is sent to or stored on our servers. While we strive for accuracy, we make no warranties about the completeness or reliability of results.