Open LLM VRAM dataset
VRAM estimates for 69 open models, generated at build time by the same functions as the calculator, the model pages and the report.
- modelvram-vram.csv (one row per model)
- modelvram-vram.json (the same rows, plus column notes, units and assumptions)
Version 1.2026-10-20; model data last checked .
Quotable figures computed from these rows, such as how many models fit a 24 GB GPU, are on the LLM VRAM statistics page.
The largest model per memory tier, the KV cache ranking of all 69 models and an embeddable chart are in the 2026 Local LLM VRAM Report.
Each model also has a static VRAM badge for its Hugging Face README.
License and attribution
The data is released under CC BY 4.0: copy, adapt and use it commercially, with attribution:
Data: ModelVRAM (modelvram.com), linking to https://modelvram.com/
Units
Columns ending in _gib are in GiB (1024³ bytes); the site writes that as GB. Parameters are counts, contexts are tokens and kv_bytes_per_token_fp16 is bytes. Values a model does not have, such as totals past its longest context, are empty.
Method
Weight memory is total parameters × bits per parameter ÷ 8 (Q4_K_M 4.84, Q8_0 8.5, FP16 / BF16 16 bits; mixture-of-experts models count every expert). The KV cache follows each model's attention layout, at FP16 for one request: full attention grows with context, sliding-window layers keep only their window, MLA stores a compressed latent, and linear or state-space layers add no context-growing cache. Total = weights + KV cache + 0.5 GB + 10% of weights and cache. The "fits" columns say whether that total at 32,768 tokens fits one GPU of the given size. Contexts are 8,192, 32,768, 131,072 tokens. gpt-oss is at each column's bits per parameter too; it is published in MXFP4 and its GGUF quants keep the experts in MXFP4, so its real files are all about the MXFP4 size, and the "Can I run", GPU and VRAM tier pages list it at MXFP4 only, with at least 0.5 GB left free.
These are estimates, not measured peaks; vLLM reservations, CUDA graphs and vision encoders can need more. See the KV cache explanation, and predicted vs measured for how they compare with public measurements. Measured a model yourself? Submit a measurement.
Sources
Model shapes and parameter counts come from Hugging Face configs and weight metadata, or Ollama GGUF headers and tensors; Nimble uses the merged full model, not the LoRA adapter alone, read on the date in source_checked. The calculation code and tests are on GitHub; the data is also on Hugging Face Datasets.
JSON API
The same numbers are served as static JSON, built with the site: no key, any origin (Access-Control-Allow-Origin: *), CC BY 4.0 like the files above. Every response carries version (the data version above), a checked date, license and source. It covers the 69 models and 73 GPU setups on the site; for any other Hugging Face model, use the calculator.
/api/v1/models.json: every model: parameters, architecture, KV bytes per token at f16 and q8_0, weights per precision (GiB) and links; builds lists low-bit builds published only in their own formats (Ternary Bonsai 2 27B) with exact file bytes/api/v1/models/<id>.json: one model: totals per precision at 4K, 8K, 32K, 128K and its longest context with f16 and q8_0 KV, checked GGUF repos, the smallest setup and a "Can I run" verdict per single GPU/api/v1/gpus.json: every GPU and multi-GPU setup/api/v1/gpus/<id>.json: the models one setup holds (best precision at 8K, total, longest context, speed) and the ones it does not
curl -s https://modelvram.com/api/v1/models/qwen3.6-35b-a3b.json | jq '.precisions[] | select(.id=="q4_k_m") | .totals_gib.f16'
curl -s https://modelvram.com/api/v1/gpus/rtx-4090.json | jq '.fits[] | {name, best_precision, total_gib}' IDs are the page slugs, such as qwen3.6-35b-a3b, rtx-4090 or 2x-rtx-3090. A site overview for AI assistants is at /llms.txt, and the calculators' URL parameters are in /agent.md. Within v1, fields may be added but never removed or changed in meaning.
Columns
| Column | Unit | Description |
|---|---|---|
model | Model name as shown on the site | |
hf_id | Hugging Face repository ID | |
hf_url | URL | Source repository on Hugging Face |
model_url | URL | ModelVRAM page with the full calculations |
source_checked | date | When model metadata and weight sizes were read |
total_params | parameters | All parameters, every expert included |
active_params | parameters | Parameters used per token (MoE); empty for dense models |
experts | count | Routed experts (MoE); empty for dense models |
layers | count | Transformer layers |
max_context | tokens | Longest context in the model config |
kv_bytes_per_token_fp16 | bytes | KV cache added per token of one request at FP16, once sliding windows are full |
attention_layout | Layers that decide the KV cache | |
q4_k_m_weights_gib | GiB | Weight memory at Q4_K_M |
q8_0_weights_gib | GiB | Weight memory at Q8_0 |
fp16_weights_gib | GiB | Weight memory at FP16 / BF16 |
q4_k_m_8k_total_gib | GiB | Total VRAM at Q4_K_M, 8,192 tokens: weights + KV cache + overhead; empty past max_context |
q4_k_m_32k_total_gib | GiB | Total VRAM at Q4_K_M, 32,768 tokens: weights + KV cache + overhead; empty past max_context |
q4_k_m_128k_total_gib | GiB | Total VRAM at Q4_K_M, 131,072 tokens: weights + KV cache + overhead; empty past max_context |
q8_0_8k_total_gib | GiB | Total VRAM at Q8_0, 8,192 tokens: weights + KV cache + overhead; empty past max_context |
q8_0_32k_total_gib | GiB | Total VRAM at Q8_0, 32,768 tokens: weights + KV cache + overhead; empty past max_context |
q8_0_128k_total_gib | GiB | Total VRAM at Q8_0, 131,072 tokens: weights + KV cache + overhead; empty past max_context |
fp16_8k_total_gib | GiB | Total VRAM at FP16 / BF16, 8,192 tokens: weights + KV cache + overhead; empty past max_context |
fp16_32k_total_gib | GiB | Total VRAM at FP16 / BF16, 32,768 tokens: weights + KV cache + overhead; empty past max_context |
fp16_128k_total_gib | GiB | Total VRAM at FP16 / BF16, 131,072 tokens: weights + KV cache + overhead; empty past max_context |
kv_cache_8k_gib | GiB | FP16 KV cache of one request at 8,192 tokens |
kv_cache_32k_gib | GiB | FP16 KV cache of one request at 32,768 tokens |
kv_cache_128k_gib | GiB | FP16 KV cache of one request at 131,072 tokens |
q4_k_m_32k_fits_16gib | boolean | Q4_K_M with 32,768 tokens fits on one 16 GB GPU |
q4_k_m_32k_fits_24gib | boolean | Q4_K_M with 32,768 tokens fits on one 24 GB GPU |
q4_k_m_32k_fits_32gib | boolean | Q4_K_M with 32,768 tokens fits on one 32 GB GPU |
q4_k_m_32k_fits_48gib | boolean | Q4_K_M with 32,768 tokens fits on one 48 GB GPU |
q4_k_m_32k_fits_80gib | boolean | Q4_K_M with 32,768 tokens fits on one 80 GB GPU |
q8_0_32k_fits_16gib | boolean | Q8_0 with 32,768 tokens fits on one 16 GB GPU |
q8_0_32k_fits_24gib | boolean | Q8_0 with 32,768 tokens fits on one 24 GB GPU |
q8_0_32k_fits_32gib | boolean | Q8_0 with 32,768 tokens fits on one 32 GB GPU |
q8_0_32k_fits_48gib | boolean | Q8_0 with 32,768 tokens fits on one 48 GB GPU |
q8_0_32k_fits_80gib | boolean | Q8_0 with 32,768 tokens fits on one 80 GB GPU |
fp16_32k_fits_16gib | boolean | FP16 / BF16 with 32,768 tokens fits on one 16 GB GPU |
fp16_32k_fits_24gib | boolean | FP16 / BF16 with 32,768 tokens fits on one 24 GB GPU |
fp16_32k_fits_32gib | boolean | FP16 / BF16 with 32,768 tokens fits on one 32 GB GPU |
fp16_32k_fits_48gib | boolean | FP16 / BF16 with 32,768 tokens fits on one 48 GB GPU |
fp16_32k_fits_80gib | boolean | FP16 / BF16 with 32,768 tokens fits on one 80 GB GPU |
note | Known cache optimization the estimate leaves out |