LLM VRAM Calculator

Use VRAM (GB) = parameters (B) × bytes_per_param × 1.2 + 2 GB for the OS. Bytes per parameter: FP16 = 2, Q8 = 1, Q5_K_M = 0.625, Q4_K_M = 0.5. An 8B model at Q4_K_M is 8 × 0.5 × 1.2 + 2 = 6.8 GB at a short context. Longer context adds KV cache on top. The method is on the methodology page.

7 × 0.5 × 1.2 + 2 = 6.2 GB at Q4_K_M, 7 × 1.0 × 1.2 + 2 = 10.4 GB at Q8, and 7 × 2.0 × 1.2 + 2 = 18.8 GB at FP16. An 8 GB card fits Q4_K_M with a short context. Q8 needs 12 GB. Pull it with ollama pull qwen3:8b when you want that size.

32 × 0.5 × 1.2 + 2 = 21.2 GB. A 24 GB GPU such as an RTX 3090 or RTX 4090 fits it at a short context. Q8 is 32 × 1.2 + 2 = 40.4 GB and does not fit one 24 GB card. The matching Ollama tag is qwen3:32b.

70 × 0.5 × 1.2 + 2 = 44 GB at Q4_K_M. That does not fit a 32 GB RTX 5090. A 64 GB or larger unified-memory Mac, or two 24 GB GPUs, can hold the weights. ollama pull llama3.3:70b is the tag, not a promise that one consumer GPU can run it.

No. The 1.2 factor plus the 2 GB OS reserve covers activations and a short context, about 4K tokens. This calculator adds an extra KV-cache allowance for 16K, 32K, and 128K. Those extras are approximate and depend on layer count and KV heads. The methodology page shows the per-token KV formula.

13 × 0.5 × 1.2 + 2 = 9.8 GB at Q4_K_M. An 8 GB card is short. A 12 GB card such as an Intel Arc B580 or an RTX 4070 fits that Q4_K_M estimate. qwen3:14b is the nearby Ollama tag and needs 10.4 GB.

Pull a size tag, not a VRAM setting. If a 14B Q4_K_M fits, run ollama pull qwen3:14b and then ollama run qwen3:14b. In LM Studio, download the Q4_K_M GGUF of the same size. If ollama ps shows GPU under 100%, the weights spilled to system RAM.

Unified memory is shared with macOS. Treat usable memory as the catalog figure minus several GB for the OS, on top of the formula’s 2 GB reserve. A 24 GB Mac is not the same fit as a 24 GB discrete GPU. See the Apple Silicon guide for the wired-memory caveat.

At a short context, yes. Qwen’s card lists 27B parameters and Apache-2.0. The formula is 27 × 0.5 × 1.2 + 2 = 18.2 GB at Q4_K_M. On 3 October 2026 Unsloth’s GGUF tree showed UD-Q4_K_S at 15.4 GB, UD-Q4_K_M at 16.5 GB, and UD-Q4_K_XL at 17.6 GB. Those are file sizes, smaller than the formula because the formula adds a 1.2 runtime factor and 2 GB. Ollama’s library listed qwen3.8:27b at 18 GB the same day. Q8 is 27 × 1.2 + 2 = 34.4 GB and does not fit 24 GB. A 16 GB card does not fit the 18.2 GB estimate.

Use the total parameter count, not the active count. Google’s card lists 25.2B total and 3.8B active, Apache-2.0. All experts stay resident, so Q4_K_M is 25.2 × 0.5 × 1.2 + 2 = 17.1 GB at a short context. Unsloth’s hardware table recommends 16–18 GB at 4-bit. Their UD-Q4_K_XL file was 17.01 GB, and Google’s QAT Q4_0 GGUF was 14.44 GB. Ollama listed gemma4:26b at 16–19 GB. A 24 GB GPU is the comfortable tier. A 16 GB GPU is the edge of the file, not a comfortable context.

No. Aleph-Alpha/Kolibri-1, released 3 October 2026 under Apache-2.0, is a 78B MoE with 3.46B active parameters. The card’s memory footprint is about 78 GB of FP8 weights, and the minimum hardware it lists is two A100 80 GB, two H100s, or one H200, B200, or B300. A Q4 formula on 78.1B would be about 48.9 GB, which still misses every 24 GB and 32 GB gaming card, and that formula is not a published consumer quant. Treat Kolibri as a datacenter or API serve.

Qwen3.6-27B is a 27B dense model under Apache-2.0. Q4 is 27 × 0.5 × 1.2 + 2 = 18.2 GB. On 5 October 2026 Unsloth’s Q4_K_M file was 16.82 GB and UD-Q4_K_XL was 17.61 GB. Ollama listed qwen3.6:27b at 18–19 GB. A 24 GB GPU fits that Q4. A 16 GB GPU does not. Qwen3.6-35B-A3B is 35B total and 3B active. Count the total: 35 × 0.5 × 1.2 + 2 = 23.0 GB. Unsloth’s UD-Q4_K_M file was 22.13 GB, and Ollama listed qwen3.6:35b at 23–24 GB. A 24 GB card can load that file with almost no extra context. A 32 GB card is the comfortable discrete tier. Do not pull qwen3.6:latest on 16 GB: that tag was the 35B size.

JetBrains lists Mellum2.1 Thinking as 12B total and 2.5B active, Apache-2.0. Count the total: 12 × 0.5 × 1.2 + 2 = 9.2 GB at Q4. On 11 October 2026 the community Q4_K_M file was 8,819,358,432 bytes (8.82 GB). A 12 GB GPU fits the formula. An 8 GB GPU does not. Q8 is 16.4 GB and misses a 16 GB card. The official card still says Ollama and LM Studio builds are coming soon, so there is no ollama run tag on this page. Weights: https://huggingface.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking

Count 125B plus the 51B n-gram table plus the 4B MTP head, which is 180B. Q4 is 180 × 0.5 × 1.2 + 2 = 110 GB. On 9 October 2026 Unsloth’s UD-Q4_K_XL shards summed to 111.33 GB and UD-IQ4_XS summed to 93.68 GB. A 32 GB RTX 5090 does not fit either file. A 128 GB unified box fits UD-IQ4_XS with room and holds UD-Q4_K_XL only with a short context. Qwen Cloud’s Qwen3.8-Flash API is not this file. The Ollama tag qwen3.8:27b is the 27B model, not Flash-Next.

Liquid’s card lists 3.12B total parameters. The SigLIP2 vision encoder is inside that total, so do not add 400M on top. Q4 is 3.12 × 0.5 × 1.2 + 2 = 3.9 GB. Q8 is 3.12 × 1.2 + 2 = 5.7 GB. BF16 is 3.12 × 2.4 + 2 = 9.5 GB. On 12 October 2026 the GGUF repo Content-Length was Q4_K_M 1,674,456,672 bytes, Q8_0 2,874,781,280 bytes, and BF16 5,403,160,160 bytes. Those files were not re-fetched on 13 October. They are smaller than the formula because the formula adds 1.2 and 2 GB. The published command is llama-server -hf LiquidAI/d1-3B-GGUF:Q8_0, then POST /v1/systemone. There is no official ollama run tag. d1 is a decision model, not a chat model. d1-omni-600M is 587M total, and Q4 is 0.587 × 0.5 × 1.2 + 2 = 2.4 GB. Weights: https://huggingface.co/LiquidAI/d1-3B

IBM’s cards list 3B, 8B, and 30B, Apache-2.0. Official GGUF sizes on 5 October 2026: granite-4.2-3b Q4_K_M 2.24 GB, granite-4.2-8b Q4_K_M 5.35 GB and Q8_0 9.35 GB, granite-4.2-30b Q4_K_M 17.72 GB. Our VRAM formula is higher because it adds the 1.2 factor and 2 GB: 3.8 GB, 6.8 GB, and 20.0 GB at Q4. An 8 GB card fits the 3B and the 8B Q4 file. It does not fit the 8B Q8 file. A 16 GB card does not fit the 30B Q4 file. A 24 GB card does. Ollama tags are granite4.2:3b (2.2 GB), granite4.2:8b (5.3 GB), and granite4.2:30b (18 GB).

Count 24.0B total, not the 2.45B active count. Q4 is 24 × 0.5 × 1.2 + 2 = 16.4 GB and misses a 16 GB GPU. Q5 is 20.0 GB and fits 24 GB. Q8 is 24 × 1.2 + 2 = 30.8 GB and fits 32 GB with about 1 GB left. BF16 is 59.6 GB and fits 128 GB unified. On 13 October 2026 the Q4_K_M file was 14,526,258,464 bytes and the Q8_0 file was 25,493,732,640 bytes. The command is llama-cli -hf LiquidAI/LFM2.5-24B-A2B-GGUF. There is no ollama run tag. ollama run lfm2.5 is the 8B model, whose Q4 is 7.0 GB. Weights: https://huggingface.co/LiquidAI/LFM2.5-24B-A2B

The safetensors count is 2,516,756,480 parameters (2.52B). The name is 2B. It is a code-search agent, not a general chat model. Q4 is 2.51675648 × 0.5 × 1.2 + 2 = 3.5 GB. Q8 is 2.51675648 × 1.2 + 2 = 5.0 GB. Both fit an 8 GB GPU. BF16 is 8.0 GB and is the edge of 8 GB. On 13 October 2026 Q8_0 was 2,679,710,912 bytes. The README calls the default Q8 approximately 2.5 GB; quote the byte count. The install command is colgrep --install-agent. There is no ollama run tag. Weights: https://huggingface.co/lightonai/colgrep-agent-2B

Count the total, not the 23.4B active count. The BF16 safetensors total on 17 October 2026 was 501,267,612,352 parameters (501.3B). Q4 is 501.3 × 0.5 × 1.2 + 2 = 302.8 GB. That misses 128 GB and 256 GB. Unsloth’s UD-Q4_K_XL files summed to 306,533,349,412 bytes (306.5 GB). The smallest Unsloth file, UD-IQ1_S, summed to 103,059,961,380 bytes (103.1 GB): it can sit on 128 GB and does not fit 96 GB. NVFP4 safetensors summed to 293,634,877,368 bytes (273.5 GiB) and are a Blackwell checkpoint, not a Mac file. ollama run beam is the cloud tag beam:cloud, not a local download. Weights: https://huggingface.co/reflection/Beam-1.0-501B-A23B

These are Apache-2.0 bundles of a Granite 4.1 base, two adapters, and a speech encoder, released 14 October 2026. Safetensors totals that day: 8,776,400,896 (8.78B) and 29,989,814,272 (29.99B). Q4 is 7.3 GB and 20.0 GB. The 8B fits an 8 GB GPU. Its Q8 is 12.5 GB and misses 12 GB. The 30B misses 16 GB and fits 24 GB. There is no ollama run granite-switch tag. ollama run granite4.1:8b is the older base model, not this bundle. Weights: https://huggingface.co/ibm-granite/granite-switch-4.1-8b and https://huggingface.co/ibm-granite/granite-switch-4.1-30b

The BF16 card lists 3,975,229,440 parameters (3.98B). Q4 is 3.98 × 0.5 × 1.2 + 2 = 4.4 GB. Q8 is 6.8 GB. Both fit 8 GB. The published Q5_K_M GGUF was 3,673,288,864 bytes on 17 October 2026. There is no Ollama tag. The card’s local path is a llama.cpp fork, not ollama run. Weights: https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-GGUF