How much VRAM do I need for a 16B–21B model?

About 15.6 GB at Q4_K_M, 26.6 GB at Q8_0 and 49.1 GB at FP16, at a ~4k context. That is for gpt-oss-20b (20.9B parameters): weights plus about 15% for the KV cache and 1 GB for the runtime.

How much VRAM does a 16B–21B model need for 32k context?

About 16.3 GB at Q4_K_M with an FP16 KV cache, or 15.9 GB with a Q8 cache, based on gpt-oss-20b's published config. Models with different attention layouts need more or less.

What is the cheapest GPU for a 16B–21B model?

Of the GPUs we track, the Radeon RX 7900 XT (20 GB, $899 list price) is the cheapest that runs it fully on the GPU at Q4.

Can a Mac run a 16B–21B model?

Yes. The cheapest Mac or unified-memory mini PC we track that runs it at Q4 is the Mac mini M6 24GB (24 GB, about 18.0 GB usable, $1,299).

Does a MoE model of the same size need less VRAM?

No. A mixture-of-experts model still has to hold all of its weights in memory, so it needs about the same VRAM as a dense model with the same total parameters. It runs faster, because only a few experts are read per token.

How much VRAM do you need for a 16B–21B model?

Short answer: about 15.6 GB at Q4_K_M, the usual quant for local use. Below: every quant, longer contexts, the cheapest hardware that fits, and every 16B–21B open model we track.

Q4_K_M: 15.6 GB 12.7 GB weights
Q8_0: 26.6 GB 22.2 GB weights
FP16: 49.1 GB 41.8 GB weights

Worked out for gpt-oss-20b (20.9B parameters, from the Hugging Face API). Formula: weights × 1.15 + 1 GB at a ~4k context (how we calculate).

Cheapest GPU that fits at Q4: Radeon RX 7900 XT (20 GB, $899). Cheapest Mac or mini PC: Mac mini M6 24GB ($1,299).

Cheapest hardware for each quant

Only devices with a sourced price. GPUs at list price (MSRP); street prices may be higher.

Cheapest GPU and Mac or mini PC for a 16B–21B model at each quant
Setting	Memory needed	Cheapest GPU	Cheapest Mac / mini PC
Q4_K_M, ~4k context	15.6 GB	Radeon RX 7900 XT 20 GB usable · $899	Mac mini M6 24GB 18 GB usable · $1,299
Q8_0, ~4k context	26.6 GB	Intel Arc Pro B70 32GB 32 GB usable · $949	Mac mini M5 Pro 64GB 48 GB usable · $3,199
FP16, ~4k context	49.1 GB	NVIDIA RTX Pro 6000 Blackwell 96GB 96 GB usable · $8,565	Framework Desktop (Ryzen AI Max+ 395) 128GB 96 GB usable · $3,449
Q4_K_M, 32k context	16.3 GB	Radeon RX 7900 XT 20 GB usable · $899	Mac mini M6 32GB 24 GB usable · $1,799

Memory by context length

Each extra token of context adds to the KV cache. Numbers for gpt-oss-20b, from its published config; a Q8 KV cache halves the cache part.

Memory needed for a 16B–21B model by context length
Context	Q4_K_M, FP16 cache	Q4_K_M, Q8 cache	Q8_0, FP16 cache
4k tokens	15.6 GB	15.6 GB	26.6 GB
8k tokens	15.7 GB	15.6 GB	26.7 GB
16k tokens	15.9 GB	15.7 GB	26.9 GB
32k tokens	16.3 GB	15.9 GB	27.3 GB
64k tokens	17.1 GB	16.3 GB	28.1 GB
128k tokens	18.7 GB	17.1 GB	29.7 GB

Hardware that runs it

Bar: Q4 memory needed vs. usable memory. Speed is estimated decode at Q4 (how we estimate); MoE shows an upper bound only.FitsTight

Tracked hardware that runs a 16B–21B model at Q4
Hardware	Q4 need vs. usable	Q4 / Q8 / FP16	Est. tok/s
Radeon RX 7600 XTGPU · $329	15.6 / 16.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	up to 112
Radeon RX 9060 XT 16GBGPU · $349	15.6 / 16.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	up to 125
Intel Arc A770 16GBGPU · $349	15.6 / 16.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	up to 218
GeForce RTX 5060 Ti 16GBGPU · $429	15.6 / 16.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	up to 174
GeForce RTX 4060 Ti 16GBGPU · $499	15.6 / 16.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	up to 112
Radeon RX 7800 XTGPU · $499	15.6 / 16.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	up to 243
Radeon RX 7900 GREGPU · $549	15.6 / 16.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	up to 224
Radeon RX 9070GPU · $549	15.6 / 16.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	up to 249
Radeon RX 6800GPU · $579	15.6 / 16.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	up to 199
Radeon RX 9070 XTGPU · $599	15.6 / 16.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	up to 249
Radeon RX 6800 XTGPU · $649	15.6 / 16.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	up to 199
NVIDIA RTX 2000 Ada 16GBGPU · $649	15.6 / 16.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	up to 87
GeForce RTX 5070 TiGPU · $749	15.6 / 16.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	up to 349
GeForce RTX 4070 Ti SuperGPU · $799	15.6 / 16.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	up to 262
GeForce RTX 4080 SuperGPU · $999	15.6 / 16.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	up to 287
GeForce RTX 5080GPU · $999	15.6 / 16.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	up to 374
Radeon RX 6900 XTGPU · $999	15.6 / 16.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	up to 199
NVIDIA RTX A4000GPU · $1,000	15.6 / 16.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	up to 174
Radeon RX 6950 XTGPU · $1,099	15.6 / 16.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	up to 224
GeForce RTX 4080GPU · $1,199	15.6 / 16.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	up to 279
Mac mini M6 24GBunified memory · $1,299	15.6 / 18.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	up to 62
Radeon RX 7900 XTGPU · $899	15.6 / 20.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	up to 312
NVIDIA RTX 4000 Ada 20GBGPU · $1,250	15.6 / 20.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	up to 140
NVIDIA Tesla P40 24GBGPU	15.6 / 24.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	up to 135
Radeon RX 7900 XTXGPU · $999	15.6 / 24.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	up to 374
GeForce RTX 3090GPU · $1,499	15.6 / 24.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	up to 365
GeForce RTX 4090GPU · $1,599	15.6 / 24.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	up to 393
Mac mini M6 32GBunified memory · $1,799	15.6 / 24.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	up to 62
GeForce RTX 3090 TiGPU · $1,999	15.6 / 24.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	up to 393
NVIDIA RTX A5000GPU · $2,250	15.6 / 24.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	up to 299
MacBook Pro M6 32GBunified memory · $2,399	15.6 / 24.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	up to 62
NVIDIA Titan RTXGPU · $2,499	15.6 / 24.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	up to 262
NVIDIA Tesla V100 32GB (PCIe)GPU	15.6 / 32.0 GB	Q4FitsQ8FitsFP16Won’t fit	up to 351
AMD Radeon Instinct MI50 32GBGPU	15.6 / 32.0 GB	Q4FitsQ8FitsFP16Won’t fit	up to 399
AMD Radeon Instinct MI60 32GBGPU	15.6 / 32.0 GB	Q4FitsQ8FitsFP16Won’t fit	up to 399
Intel Arc Pro B70 32GBGPU · $949	15.6 / 32.0 GB	Q4FitsQ8FitsFP16Won’t fit	up to 237
GeForce RTX 5090GPU · $1,999	15.6 / 32.0 GB	Q4FitsQ8FitsFP16Won’t fit	up to 698
Mac mini M5 Pro 64GBunified memory · $3,199	15.6 / 48.0 GB	Q4FitsQ8FitsFP16Won’t fit	up to 113
NVIDIA RTX A6000GPU · $4,649	15.6 / 48.0 GB	Q4FitsQ8FitsFP16Won’t fit	up to 299
NVIDIA DGX Spark 64GBunified memory · $4,999	15.6 / 48.0 GB	Q4FitsQ8FitsFP16Won’t fit	up to 100
NVIDIA RTX 6000 Ada 48GBGPU · $6,799	15.6 / 48.0 GB	Q4FitsQ8FitsFP16Won’t fit	up to 374
Framework Desktop (Ryzen AI Max+ 395) 128GBunified memory · $3,449	15.6 / 96.0 GB	Q4FitsQ8FitsFP16Fits	up to 94
MacBook Pro M4 Max 128GBunified memory · $4,999	15.6 / 96.0 GB	Q4FitsQ8FitsFP16Fits	up to 200
MacBook Pro M5 Max 128GBunified memory · $6,699	15.6 / 96.0 GB	Q4FitsQ8FitsFP16Fits	up to 225
NVIDIA DGX Spark 128GBunified memory · $6,950	15.6 / 96.0 GB	Q4FitsQ8FitsFP16Fits	up to 100
NVIDIA RTX Pro 6000 Blackwell 96GBGPU · $8,565	15.6 / 96.0 GB	Q4FitsQ8FitsFP16Fits	up to 698
Mac Studio M5 Ultra 256GBunified memory · $9,499	15.6 / 192.0 GB	Q4FitsQ8FitsFP16Fits	up to 440
Mac Studio M5 Ultra 512GBunified memory · $22,299	15.6 / 384.0 GB	Q4FitsQ8FitsFP16Fits	up to 440

16B–21B models we track

Memory needed by each 16B–21B model we track
Model	Q4_K_M	Q8_0	FP16
LLaDA2.1-mini16.3B params · inclusionAI · MoE	12.3 GB	20.9 GB	38.4 GB
LLaDA2.0-mini16.3B params · inclusionAI · MoE	12.3 GB	20.9 GB	38.4 GB
Ling-mini-2.016.3B params · inclusionAI · MoE	12.3 GB	20.9 GB	38.4 GB
Ring-mini-2.016.3B params · inclusionAI · MoE	12.3 GB	20.9 GB	38.4 GB
LLaDA2.2-mini16.3B params · inclusionAI · MoE	12.3 GB	20.9 GB	38.4 GB
LLaDA2.0-mini-preview16.3B params · inclusionAI · MoE	12.3 GB	20.9 GB	38.4 GB
Kimi-VL-A3B-Instruct16.4B params · moonshotai · MoE	12.4 GB	21.0 GB	38.7 GB
Kimi-VL-A3B-Thinking16.4B params · moonshotai · MoE	12.4 GB	21.0 GB	38.7 GB
Kimi-VL-A3B-Thinking-250616.4B params · moonshotai · MoE	12.4 GB	21.0 GB	38.7 GB
Ring-mini-linear-2.016.4B params · inclusionAI · MoE	12.5 GB	21.1 GB	38.8 GB
Ling-lite16.8B params · inclusionAI · MoE	12.7 GB	21.5 GB	39.6 GB
Ling-lite-1.5-250716.8B params · inclusionAI · MoE	12.7 GB	21.5 GB	39.6 GB
Ling-lite-1.516.8B params · inclusionAI · MoE	12.7 GB	21.5 GB	39.6 GB
Ring-lite16.8B params · inclusionAI · MoE	12.7 GB	21.5 GB	39.6 GB
Ling-Coder-lite16.8B params · inclusionAI · MoE	12.7 GB	21.5 GB	39.6 GB
Ring-lite-distill-preview16.8B params · inclusionAI · MoE	12.7 GB	21.5 GB	39.6 GB
Groot-19B-A3B-26102319.4B params · GrootTeam · MoE	14.5 GB	24.7 GB	45.7 GB
internlm2_5-20b-chat19.9B params · internlm	14.8 GB	25.3 GB	46.7 GB
maple-preview20.2B params · deepgrove · MoE	15.1 GB	25.7 GB	47.5 GB
reka-flash-320.9B params · RekaAI	15.6 GB	26.5 GB	49.1 GB
gpt-oss-20b20.9B params · openai · MoE	15.6 GB	26.6 GB	49.1 GB
ERNIE-4.5-21B-A3B-Thinking21.8B params · baidu · MoE	16.2 GB	27.7 GB	51.2 GB
ERNIE-4.5-21B-A3B-PT21.9B params · baidu · MoE	16.3 GB	27.8 GB	51.5 GB

Have a different amount of memory? Try the VRAM calculator or see best LLMs by VRAM size.

Quick answers

How much VRAM do I need for a 16B–21B model?: About 15.6 GB at Q4_K_M, 26.6 GB at Q8_0 and 49.1 GB at FP16, at a ~4k context. That is for gpt-oss-20b (20.9B parameters): weights plus about 15% for the KV cache and 1 GB for the runtime.
How much VRAM does a 16B–21B model need for 32k context?: About 16.3 GB at Q4_K_M with an FP16 KV cache, or 15.9 GB with a Q8 cache, based on gpt-oss-20b's published config. Models with different attention layouts need more or less.
What is the cheapest GPU for a 16B–21B model?: Of the GPUs we track, the Radeon RX 7900 XT (20 GB, $899 list price) is the cheapest that runs it fully on the GPU at Q4.
Can a Mac run a 16B–21B model?: Yes. The cheapest Mac or unified-memory mini PC we track that runs it at Q4 is the Mac mini M6 24GB (24 GB, about 18.0 GB usable, $1,299).
Does a MoE model of the same size need less VRAM?: No. A mixture-of-experts model still has to hold all of its weights in memory, so it needs about the same VRAM as a dense model with the same total parameters. It runs faster, because only a few experts are read per token.

Other sizes: 1B–2B · 2B–3B · 4B–6B · 7B–8B · 12B–14B · 24B–27B · 30B–32B · 40B–50B · 70B · 120B · 200B–300B · 300B–480B · 500B–800B · 1T+ · All sizes · Ollama VRAM by tag · GGUF sizes by quant