Which local LLMs can your PC run?
Pick your hardware to see every model’s verdict, recommended quant and estimated speed band — or start from a model to see which GPUs run it.
A model that isn't listed
Paste a Hugging Face model id or link (for example Qwen/Qwen3-4B) and we'll run the same calculation for it. Your browser asks Hugging Face for the model details directly.
Enter the model's numbers yourself
Type the numbers from the model card or its config.json and we'll calculate the model the same way. This runs in your browser, and nothing is requested from Hugging Face.
In billions. Llama 3.1 8B, for example, is 8.03.
num_hidden_layers in config.json.
num_key_value_heads, or num_attention_heads if it isn't there. If config.json has num_global_key_value_heads, use that instead.
head_dim, or hidden_size divided by num_attention_heads if it isn't there. If config.json has global_head_dim, use that instead.
Only for models that mix in sliding-window or linear-attention layers: the number of full_attention entries in layer_types. Leave it empty to count every layer.
Pick the step closest to max_position_embeddings in config.json, or Not sure.
Results
Pick a GPU or chip above to see verdicts.
What runs on popular hardware
For six common setups, here is the model the calculator lists first in each size class. Pick that hardware above and turn on only that size filter, and you get the same card. Popular models come first on several setups, so you can compare how the same model runs on each one. Each setup also shows the largest model that fits entirely on its GPU.
Each graphics card is assumed to be the only GPU in the PC and to drive the display, with 32 GB of DDR5-5600 system RAM and Windows. Every setup uses 8k context and an f16 KV cache, the same as our GPU pages.
GeForce RTX 4060
8 GB VRAM · 32 GB RAM
30 of 58 models run great or well.
- Under 10B
- Qwen3.5 9BRuns greatthinking modeQ4_K_M · Full GPUest. 31.0 tok/s27.3–34.7calibrated estimate ±12%
- 10B to under 25B
- gpt-oss-20bRuns wellthinking modeMXFP4 · MoE experts in RAMest. 36.9 tok/s25.9–48.0theoretical estimate ±30%
- 25B to under 50B
- Qwen3.5 35B-A3BRuns wellthinking modeQ4_K_M · MoE experts in RAMest. 39.3 tok/s27.5–51.1theoretical estimate ±30%
- 50B and up
- Won't runall 10 models this size1 of them runs with a lower quant.Kolibri 1 with 64 GB of system RAM:Runs well
Largest model that fits entirely on the GPU
Gemma 4 12BRuns wellthinking modeIQ4_XS · Full GPUest. 29.2 tok/s25.7–32.7calibrated estimate ±12%GeForce RTX 3060 12GB
12 GB VRAM · 32 GB RAM
35 of 58 models run great or well.
- Under 10B
- Qwen3.5 9BRuns greatthinking modeQ4_K_M · Full GPU51.9 tok/s8k estimate 41.1 tok/s (36.1–46.0, calibrated estimate ±12%)measured (1 run, 32k context)
- 10B to under 25B
- gpt-oss-20bRuns wellthinking modeMXFP4 · MoE experts in RAMest. 63.9 tok/s44.8–83.1theoretical estimate ±30%
- 25B to under 50B
- Qwen3.5 35B-A3BRuns wellthinking modeQ4_K_M · MoE experts in RAMest. 49.4 tok/s34.6–64.2theoretical estimate ±30%
- 50B and up
- Won't runall 10 models this size2 of them run with a lower quant.Kolibri 1 with 64 GB of system RAM:Runs well
Largest model that fits entirely on the GPU
Kanana 1.5 15.7B-A3BRuns greatIQ4_XS · Full GPUest. 79.6 tok/s63.7–95.5calibrated estimate ±20%GeForce RTX 5060 Ti 16GB
16 GB VRAM · 32 GB RAM
37 of 58 models run great or well.
- Under 10B
- Qwen3.5 9BRuns greatthinking modeQ4_K_M · Full GPUest. 51.1 tok/s45.0–57.2calibrated estimate ±12%
- 10B to under 25B
- gpt-oss-20bRuns greatthinking modeMXFP4 · Full GPU111.6 tok/s8k estimate 104.3 tok/s (83.5–125.2, calibrated estimate ±20%)measured (1 run, 2k context)
- 25B to under 50B
- Qwen3.5 35B-A3BRuns wellthinking modeQ4_K_M · MoE experts in RAMest. 62.4 tok/s43.7–81.1theoretical estimate ±30%
- 50B and up
- Won't runall 10 models this size3 of them run with a lower quant.Kolibri 1 with 64 GB of system RAM:Runs well
Largest model that fits entirely on the GPU
Muse Glimmer 30BRuns wellthinking modeIQ4_XS · Full GPUest. 20.2 tok/s17.7–22.6calibrated estimate ±12%GeForce RTX 3090
24 GB VRAM · 32 GB RAM
48 of 58 models run great or well.
- Under 10B
- Qwen3.5 9BRuns greatthinking modeQ4_K_M · Full GPUest. 106.7 tok/s93.9–119.5calibrated estimate ±12%
- 10B to under 25B
- gpt-oss-20bRuns greatthinking modeMXFP4 · Full GPU162.0 tok/s8k estimate 201.1 tok/s (160.9–241.3, calibrated estimate ±20%)measured (1 run, 2k context)
- 25B to under 50B
- Qwen3.5 35B-A3BRuns greatthinking modeQ4_K_M · Full GPUest. 148.8 tok/s119.1–178.6calibrated estimate ±20%
- 50B and up
- Won't runall 10 models this size4 of them run with a lower quant.Kolibri 1 with 64 GB of system RAM:Runs well
Largest model that fits entirely on the GPU
Qwen3.5 35B-A3BRuns greatthinking modeQ4_K_M · Full GPUest. 148.8 tok/s119.1–178.6calibrated estimate ±20%Apple M4 Pro
64 GB unified memory · macOS
38 of 58 models run great or well.
- Under 10B
- Qwen3.5 9BRuns greatthinking modeQ4_K_M · Unified memory37.0 tok/s8k estimate 33.8 tok/s (27.0–40.6, calibrated estimate ±20%)measured (1 run, 4k context)
- 10B to under 25B
- gpt-oss-20bRuns greatthinking modeMXFP4 · Unified memoryest. 69.9 tok/s55.9–83.9calibrated estimate ±20%
- 25B to under 50B
- Qwen3.5 35B-A3BRuns greatthinking modeQ4_K_M · Unified memoryest. 60.3 tok/s48.3–72.4calibrated estimate ±20%
- 50B and up
- Kolibri 1Runs slowlythinking modeQ4_K_M · CPU onlyest. 49.5 tok/s34.6–64.3theoretical estimate ±30%
Largest model that fits entirely on the GPU
Kolibri 1Runs greatthinking modeIQ4_XS · Unified memoryest. 52.1 tok/s41.7–62.5calibrated estimate ±20%Ryzen AI Max+ 395 (Strix Halo)
128 GB unified memory · Linux
46 of 58 models run great or well.
- Under 10B
- Qwen3.5 9BRuns greatthinking modeQ4_K_M · Unified memory37.2 tok/s8k estimate 34.2 tok/s (27.4–41.0, calibrated estimate ±20%)measured (1 run, 4k context)
- 10B to under 25B
- gpt-oss-20bRuns greatthinking modeMXFP4 · Unified memoryest. 68.6 tok/s54.8–82.3calibrated estimate ±20%
- 25B to under 50B
- Qwen3.5 35B-A3BRuns greatthinking modeQ4_K_M · Unified memory64.8 tok/s8k estimate 56.0 tok/s (44.8–67.2, calibrated estimate ±20%)measured (1 run, 512 context)
- 50B and up
- gpt-oss-120bRuns greatthinking modeMXFP4 · Unified memory50.2 tok/s8k estimate 48.5 tok/s (38.8–58.1, calibrated estimate ±20%)measured (1 run, 512 context)
Largest model that fits entirely on the GPU
Ling-3.0 FlashRuns greatthinking modeQ4_K_M · Unified memoryest. 48.4 tok/s33.9–62.9theoretical estimate ±30%
Models with a thinking mode spend extra tokens thinking, so their speed thresholds are 1.5× stricter (30 / 12 / 3 tok/s). With thinking turned off, a model may do one grade better.
How to read the results
Pick your GPU, Mac chip or CPU above, or enter its specs by hand. Every model in the catalog then gets a card. Models that run great or well come first: the largest size class first, and the more popular model first within a class.
- Verdict
- How usable the model is on that hardware, judged on the Q4_K_M file. A model without that file is judged on its closest file (MXFP4 for gpt-oss). “Runs great” means the answer comes faster than you can read it, and “Won't run” means the model doesn't fit or is too slow to use. The thresholds for all four levels are below.
- Recommended quant
- The GGUF file to download. It's the file the verdict is based on, unless that one won't run. Then the card switches to the file that runs best.
- Estimated speed
- How fast the answer is written, in tokens per second. It comes as a range with a confidence label, and the FAQ below explains the labels.
- Memory
- The first number is what the model uses: its weights, its context and working buffers, plus what the OS keeps for the display when the card drives one. The second number is all of the VRAM on a graphics card. On a Mac or another unified-memory PC, it is the share we expect the GPU to use: 70% of memory, or on a DGX Spark what is left after Linux and the desktop. That's why a model that fits in the installed memory can still end up on the CPU only. When part of the model stays in system RAM, that amount is shown as well.
- Runs as
- Where the model is loaded.
- Full GPU: the whole model is in the graphics card's memory (VRAM).
- Unified memory: the whole model is on the GPU of a Mac or another unified-memory PC.
- Partial offload: the model is split between the GPU and system RAM.
- MoE experts in RAM: the MoE experts that don't fit in VRAM stay in system RAM.
- CPU only: the CPU does all the work. On a Mac or another unified-memory PC, this is what we assume when the model is bigger than the GPU's share of memory.
- Context headroom
- The longest context that keeps the same verdict and the same placement.
- Details
- The page for that model on that GPU, with every quant, a context-length table and the commands for Ollama and llama.cpp.
How to read the verdicts
- Runs great
- Fully on the GPU at 20 tok/s or more
- Runs well
- Fully on the GPU at 8–20 tok/s; MoE experts in system RAM at 20 tok/s or more; or at least 90% on the GPU at 8 tok/s or more
- Runs slowly
- 2–8 tok/s; CPU-only; less than 90% on the GPU; or MoE experts in system RAM below 20 tok/s
- Won't run
- Does not fit, or under 2 tok/s
The calculator starts from the same conditions as the examples above. Macs and other unified-memory machines start at their largest memory option. Change anything that differs from your machine.
Shopping for hardware instead? Switch to “Find GPUs for a model” and pick a model to see which GPUs, Macs and CPUs run it, along with the cheapest GPU by street price that runs it great or well.
Frequently asked questions
How accurate are the speed estimates?
Every speed is a band with a confidence label: measured (a public benchmark for this exact GPU, model and quant), calibrated (a bandwidth model fitted to public benchmarks — ±12% for NVIDIA and AMD cards running a dense model fully on the GPU, ±20% otherwise) or theoretical (±30%, outside our calibration set). Drivers, backend, context and thermals move real results within that band.
Can I count two or more graphics cards?
Yes, as a beta. Pick a graphics card, then raise “Number of GPUs”. With an NVIDIA or AMD desktop card, “Add a different GPU” also lets you mix in a different card from the same maker. VRAM adds up but memory bandwidth does not, so a model that already fits on one card gets no faster. What you gain is room for a bigger model on the GPUs. When you mix cards, the slower card's share slows the whole run down. Speeds split across cards are all shown as theoretical estimates (±30%).
Is my hardware information sent anywhere?
No. Detection and every calculation run inside your browser; there is no server. A share link only carries the GPU, RAM and settings you picked as URL parameters.
Speeds are estimates from memory bandwidth, calibrated against public benchmarks, and each one comes with an error band and a confidence label. Real results vary with drivers, backend, context length and thermals.