Best LLMs to run on RTX A4000 16GB
datacenter 16 GB GDDR6 · 448 GB/s · 77 TFLOPS BF16 · updated 2026-10-23
Headline answer
Of 141 models in our catalog, 53 fit on a single RTX A4000 16GB at FP8, 63 at 4-bit (AWQ) and 29 at BF16 (8K context, batch 1). Largest that fits: Qwen3-VL 4B at BF16; GLM-4 9B at FP8; DeepSeek Coder V2 Lite at 4-bit (AWQ). Rental: $0.08/hr catalog estimate.
Change context, concurrency or framework in the calculator → RTX A4000 16GB full specs
Every model on RTX A4000 16GB
Fit band and share of VRAM per precision; N× GPUs = minimum count inside one node (tensor parallel); — = does not fit within one 8-GPU node. Speed and $/1M tokens for the highest-quality precision that fits on one GPU (BF16, else FP8, else 4-bit), single stream.
| Model | Params | BF16 | FP8 | 4-bit (AWQ) | tok/s | $ / 1M tok |
|---|---|---|---|---|---|---|
| DeepSeek Coder V2 Lite | 16.0B (2.4B active) | 4× GPUs | 2× GPUs | tight 83% | 132 4-bit | $0.17 |
| StarCoder2 15B | 16.0B | 4× GPUs | 2× GPUs | tight 87% | 30 4-bit | $0.73 |
| Phi 3 Medium | 14.0B | 4× GPUs | 2× GPUs | tight 84% | 34 4-bit | $0.66 |
| Phi-4 14B | 14.0B | 4× GPUs | 2× GPUs | tight 91% | 34 4-bit | $0.66 |
| LLaMA 2 13B 4K ctx | 13.0B | 4× GPUs | 2× GPUs | tight 92% | 34 4-bit | $0.66 |
| Mistral Nemo 12B | 12.0B | 4× GPUs | 2× GPUs | tight 85% | 39 4-bit | $0.57 |
| Nemotron-Nano-12B-v2 | 12.0B | 4× GPUs | 2× GPUs | tight 90% | 38 4-bit | $0.58 |
| StableLM 2 12B 4K ctx | 12.0B | 4× GPUs | 2× GPUs | good 79% | 39 4-bit | $0.58 |
| Falcon3 10B | 10.0B | 2× GPUs | 2× GPUs | good 74% | 45 4-bit | $0.49 |
| GLM-4 9B | 9.0B | 2× GPUs | tight 85% | optimal 68% | 27 FP8 | $0.81 |
| Gemma 2 9B | 9.0B | 2× GPUs | 2× GPUs | tight 89% | 48 4-bit | $0.46 |
| Nemotron-Nano-9B-v2 | 9.0B | 2× GPUs | tight 95% | good 77% | 27 FP8 | $0.83 |
| Yi-1.5 9B 4K ctx | 9.0B | 2× GPUs | tight 85% | optimal 62% | 27 FP8 | $0.82 |
| InternLM3 8B | 8.8B | 2× GPUs | tight 84% | optimal 66% | 28 FP8 | $0.80 |
| Qwen3-VL 8B | 8.8B | 2× GPUs | tight 89% | good 73% | 28 FP8 | $0.80 |
| MiniCPM-V 4.5 | 8.7B | 2× GPUs | tight 88% | good 73% | 28 FP8 | $0.80 |
| Ministral-3 8B | 8.5B | 2× GPUs | tight 87% | good 70% | 29 FP8 | $0.78 |
| Qwen2.5-VL 7B | 8.3B | 2× GPUs | tight 81% | optimal 65% | 29 FP8 | $0.76 |
| DeepSeek R1 Distill 8B | 8.0B | 2× GPUs | tight 83% | optimal 68% | 30 FP8 | $0.74 |
| Granite 3.3 8B | 8.0B | 2× GPUs | tight 85% | optimal 64% | 30 FP8 | $0.74 |
| LLaMA 3.1 8B | 8.0B | 2× GPUs | tight 83% | optimal 68% | 30 FP8 | $0.74 |
| Ministral 8B | 8.0B | 2× GPUs | tight 84% | optimal 69% | 30 FP8 | $0.74 |
| Qwen3 8B | 8.0B | 2× GPUs | tight 84% | good 70% | 30 FP8 | $0.74 |
| Qwen 2.5 7B | 7.6B | 2× GPUs | good 77% | optimal 63% | 32 FP8 | $0.70 |
| Falcon3 7B | 7.5B | 2× GPUs | good 79% | optimal 63% | 32 FP8 | $0.69 |
| Phi-3 Small 7B | 7.4B | 2× GPUs | good 79% | optimal 64% | 32 FP8 | $0.69 |
| InternLM 2.5 7B | 7.0B | 2× GPUs | good 77% | optimal 62% | 34 FP8 | $0.65 |
| LLaMA 2 7B 4K ctx | 7.0B | 2× GPUs | tight 84% | optimal 64% | 33 FP8 | $0.68 |
| Mistral 7B | 7.0B | 2× GPUs | good 77% | optimal 57% | 34 FP8 | $0.65 |
| OLMo-2 7B 4K ctx | 7.0B | 2× GPUs | tight 84% | optimal 69% | 33 FP8 | $0.68 |
| Qwen 2 7B | 7.0B | 2× GPUs | good 73% | optimal 61% | 34 FP8 | $0.65 |
| Qwen 2.5 Coder 7B | 7.0B | 2× GPUs | good 73% | optimal 61% | 34 FP8 | $0.65 |
| StarCoder2 7B | 7.0B | 2× GPUs | good 73% | optimal 56% | 34 FP8 | $0.65 |
| Yi-1.5 6B 4K ctx | 6.0B | 2× GPUs | optimal 65% | optimal 52% | 39 FP8 | $0.57 |
| Qwen3-VL 4B | 4.4B | tight 88% | optimal 61% | optimal 55% | 27 BF16 | $0.81 |
| Gemma 3 4B | 4.3B | tight 86% | optimal 60% | optimal 59% | 28 BF16 | $0.79 |
| Nemotron-Mini-4B 4K ctx | 4.0B | good 79% | optimal 54% | optimal 56% | 30 BF16 | $0.74 |
| Qwen3 4B | 4.0B | tight 83% | optimal 58% | optimal 53% | 30 BF16 | $0.74 |
| Ministral 3B | 3.8B | good 77% | optimal 54% | optimal 49% | 32 BF16 | $0.70 |
| Phi 3.5 Mini | 3.8B | tight 93% | optimal 70% | optimal 60% | 31 BF16 | $0.72 |
| Phi-4 Mini | 3.8B | good 80% | optimal 56% | optimal 56% | 32 BF16 | $0.70 |
| Qwen2.5-VL 3B | 3.8B | good 74% | optimal 51% | optimal 45% | 32 BF16 | $0.69 |
| Ministral-3 3B | 3.4B | good 73% | optimal 53% | optimal 49% | 35 BF16 | $0.63 |
| LLaMA 3.2 3B | 3.2B | good 71% | optimal 52% | optimal 49% | 37 BF16 | $0.61 |
| Falcon3 3B | 3.0B | optimal 68% | optimal 49% | optimal 47% | 39 BF16 | $0.57 |
| Qwen 2.5 3B | 3.0B | optimal 65% | optimal 46% | optimal 43% | 39 BF16 | $0.57 |
| Qwen 2.5 Coder 3B | 3.0B | optimal 65% | optimal 46% | optimal 43% | 39 BF16 | $0.57 |
| StarCoder2 3B | 3.0B | optimal 65% | optimal 46% | optimal 40% | 39 BF16 | $0.56 |
| Qwen3-VL 2B | 2.1B | optimal 58% | optimal 45% | optimal 44% | 52 BF16 | $0.43 |
| Granite 3.1 2B | 2.0B | optimal 55% | optimal 42% | optimal 38% | 55 BF16 | $0.40 |
| Qwen3 1.7B | 1.7B | optimal 53% | optimal 42% | optimal 43% | 63 BF16 | $0.35 |
| SmolLM2 1.7B | 1.7B | optimal 57% | optimal 46% | optimal 43% | 62 BF16 | $0.36 |
| StableLM 2 1.6B 4K ctx | 1.6B | optimal 50% | optimal 41% | optimal 40% | 65 BF16 | $0.34 |
| Qwen 2 1.5B | 1.5B | optimal 46% | optimal 36% | optimal 36% | 70 BF16 | $0.32 |
| Qwen 2.5 1.5B | 1.5B | optimal 46% | optimal 36% | optimal 36% | 70 BF16 | $0.32 |
| Qwen 2.5 Coder 1.5B | 1.5B | optimal 46% | optimal 36% | optimal 36% | 70 BF16 | $0.32 |
| LLaMA 3.2 1B | 1.2B | optimal 43% | optimal 35% | optimal 36% | 81 BF16 | $0.27 |
| Falcon3 1B 4K ctx | 1.0B | optimal 40% | optimal 34% | optimal 35% | 94 BF16 | $0.24 |
| Gemma 3 1B | 1.0B | optimal 39% | optimal 33% | optimal 35% | 95 BF16 | $0.23 |
| Qwen3 0.6B | 0.6B | optimal 39% | optimal 35% | optimal 36% | 128 BF16 | $0.17 |
| Qwen 2 0.5B | 0.5B | optimal 32% | optimal 29% | optimal 30% | 148 BF16 | $0.15 |
| Qwen 2.5 0.5B | 0.5B | optimal 32% | optimal 29% | optimal 30% | 148 BF16 | $0.15 |
| Qwen 2.5 Coder 0.5B | 0.5B | optimal 32% | optimal 29% | optimal 30% | 148 BF16 | $0.15 |
| Devstral 2 123B | 123.0B | — | — | 8× GPUs | 8 4-bit, 8× GPUs | $22.22 |
| Mistral Large 2 123B | 123.0B | — | — | 8× GPUs | 8 4-bit, 8× GPUs | $22.22 |
| Command R+ 104B | 104.0B | — | — | 8× GPUs | 9 4-bit, 8× GPUs | $18.91 |
| Qwen 2 72B | 72.0B | — | 8× GPUs | 8× GPUs | 7 FP8, 8× GPUs | $25.76 |
| Qwen 2.5 72B | 72.0B | — | 8× GPUs | 8× GPUs | 7 FP8, 8× GPUs | $25.76 |
| CodeLlama 70B | 70.0B | — | 8× GPUs | 4× GPUs | 7 FP8, 8× GPUs | $25.04 |
| DeepSeek R1 Distill 70B | 70.0B | — | 8× GPUs | 4× GPUs | 7 FP8, 8× GPUs | $25.04 |
| LLaMA 2 70B 4K ctx | 70.0B | — | 8× GPUs | 4× GPUs | 7 FP8, 8× GPUs | $25.04 |
| LLaMA 3.1 70B | 70.0B | — | 8× GPUs | 4× GPUs | 7 FP8, 8× GPUs | $25.04 |
| LLaMA 3.3 70B | 70.0B | — | 8× GPUs | 4× GPUs | 7 FP8, 8× GPUs | $25.04 |
| Llama-3.1-Nemotron-70B | 70.0B | — | 8× GPUs | 4× GPUs | 7 FP8, 8× GPUs | $25.04 |
| Nemotron-H 56B Base | 56.0B | — | 8× GPUs | 4× GPUs | 9 FP8, 8× GPUs | $20.43 |
| Llama-3.1-Nemotron-51B | 51.0B | — | 8× GPUs | 4× GPUs | 10 FP8, 8× GPUs | $18.52 |
| Llama-3.3-Nemotron Super 49B | 49.0B | — | 8× GPUs | 4× GPUs | 10 FP8, 8× GPUs | $17.96 |
| Kimi Linear 48B-A3B | 48.0B (3.0B active) | — | — | 4× GPUs | 96 4-bit, 4× GPUs | $0.93 |
| Mixtral 8x7B | 46.7B (12.9B active) | — | — | 4× GPUs | 49 4-bit, 4× GPUs | $1.81 |
| Command R 35B | 35.0B | 8× GPUs | 8× GPUs | 4× GPUs | 13 FP8, 8× GPUs | $13.37 |
| Qwen3.6 35B-A3B | 35.0B (3.0B active) | — | 4× GPUs | 4× GPUs | 72 FP8, 4× GPUs | $1.24 |
| CodeLlama 34B | 34.0B | 8× GPUs | 4× GPUs | 2× GPUs | 15 FP8, 4× GPUs | $6.09 |
| Yi-1.5 34B 4K ctx | 34.0B | 8× GPUs | 4× GPUs | 2× GPUs | 14 FP8, 4× GPUs | $6.13 |
| DeepSeek R1 Distill 32B | 32.0B | 8× GPUs | 4× GPUs | 2× GPUs | 15 FP8, 4× GPUs | $5.77 |
| Nemotron-Terminal-32B | 32.0B | 8× GPUs | 4× GPUs | 2× GPUs | 15 FP8, 4× GPUs | $5.77 |
| OLMo-2 32B 4K ctx | 32.0B | 8× GPUs | 4× GPUs | 2× GPUs | 15 FP8, 4× GPUs | $5.77 |
| Qwen 2.5 32B | 32.0B | 8× GPUs | 4× GPUs | 2× GPUs | 15 FP8, 4× GPUs | $5.77 |
| Qwen 2.5 Coder 32B | 32.0B | 8× GPUs | 4× GPUs | 2× GPUs | 15 FP8, 4× GPUs | $5.77 |
| Qwen3 32B | 32.0B | 8× GPUs | 4× GPUs | 2× GPUs | 15 FP8, 4× GPUs | $5.77 |
| Nemotron-3 Nano 30B-A3B | 31.6B (3.2B active) | — | 4× GPUs | 2× GPUs | 51 FP8, 4× GPUs | $1.75 |
| Muse Glimmer 30B | 30.0B | 8× GPUs | 4× GPUs | 2× GPUs | 16 FP8, 4× GPUs | $5.42 |
| Qwen3 30B-A3B | 30.0B (3.0B active) | — | 4× GPUs | 2× GPUs | 63 FP8, 4× GPUs | $1.41 |
| Gemma 2 27B | 27.0B | 8× GPUs | 4× GPUs | 2× GPUs | 18 FP8, 4× GPUs | $4.94 |
| Gemma 3 27B | 27.0B | 8× GPUs | 4× GPUs | 4× GPUs | 18 FP8, 4× GPUs | $4.94 |
| Qwen3.6 27B | 27.0B | 8× GPUs | 4× GPUs | 2× GPUs | 18 FP8, 4× GPUs | $4.91 |
| Devstral 24B | 24.0B | 8× GPUs | 4× GPUs | 2× GPUs | 20 FP8, 4× GPUs | $4.40 |
| Devstral Small 2 24B | 24.0B | 8× GPUs | 4× GPUs | 2× GPUs | 20 FP8, 4× GPUs | $4.40 |
| Magistral Small 24B | 24.0B | 8× GPUs | 4× GPUs | 2× GPUs | 20 FP8, 4× GPUs | $4.40 |
| Mistral Small 3.1 24B | 24.0B | 8× GPUs | 4× GPUs | 2× GPUs | 20 FP8, 4× GPUs | $4.40 |
| Codestral 22B | 22.0B | 8× GPUs | 4× GPUs | 2× GPUs | 22 FP8, 4× GPUs | $4.06 |
| GPT OSS 20B | 21.0B (3.6B active) | — | 4× GPUs | 2× GPUs | 61 FP8, 4× GPUs | $1.47 |
| InternLM 2.5 20B | 20.0B | 4× GPUs | 4× GPUs | 2× GPUs | 24 FP8, 4× GPUs | $3.70 |
| DeepSeek R1 Distill 14B | 14.8B | 4× GPUs | 2× GPUs | 2× GPUs | 23 FP8, 2× GPUs | $1.96 |
| Qwen 2.5 14B | 14.8B | 4× GPUs | 2× GPUs | 2× GPUs | 23 FP8, 2× GPUs | $1.96 |
| Qwen3 14B | 14.8B | 4× GPUs | 2× GPUs | 2× GPUs | 23 FP8, 2× GPUs | $1.95 |
| Qwen 2.5 Coder 14B | 14.0B | 4× GPUs | 2× GPUs | 2× GPUs | 24 FP8, 2× GPUs | $1.86 |
| OLMo-2 13B 4K ctx | 13.7B | 4× GPUs | 2× GPUs | 2× GPUs | 24 FP8, 2× GPUs | $1.88 |
| Gemma 3 12B | 12.0B | 4× GPUs | 2× GPUs | 2× GPUs | 27 FP8, 2× GPUs | $1.63 |
| Kimi K3 | 2800.0B (104.0B active) | — | — | — | — | — |
| DeepSeek V4-Pro | 1600.0B (49.0B active) | — | — | — | — | — |
| Kimi K2 | 1000.0B (32.0B active) | — | — | — | — | — |
| Kimi K2.5 | 1000.0B (32.0B active) | — | — | — | — | — |
| GLM-5 744B | 744.0B (40.0B active) | — | — | — | — | — |
| GLM-5.1 744B | 744.0B (40.0B active) | — | — | — | — | — |
| DeepSeek R1 | 671.0B (37.0B active) | — | — | — | — | — |
| DeepSeek R1-0528 | 671.0B (37.0B active) | — | — | — | — | — |
| DeepSeek V3 | 671.0B (37.0B active) | — | — | — | — | — |
| DeepSeek V3-0324 | 671.0B (37.0B active) | — | — | — | — | — |
| DeepSeek V3.2 | 671.0B (37.0B active) | — | — | — | — | — |
| MiniMax-M1 | 456.0B (45.9B active) | — | — | — | — | — |
| MiniMax-Text-01 | 456.0B (45.9B active) | — | — | — | — | — |
| LLaMA 3.1 405B | 405.0B | — | — | — | — | — |
| Llama 4 Maverick (400B) | 400.0B (17.0B active) | — | — | — | — | — |
| Qwen3.5 397B-A17B | 397.0B (17.0B active) | — | — | — | — | — |
| Nemotron-4 340B | 340.0B | — | — | — | — | — |
| MiMo V2-Flash | 309.8B (15.0B active) | — | — | — | — | — |
| DeepSeek V4-Flash | 284.0B (13.0B active) | — | — | — | — | — |
| Llama-3.3-Nemotron Ultra 253B | 253.0B | — | — | — | — | — |
| DeepSeek Coder V2 | 236.0B (21.0B active) | — | — | — | — | — |
| Qwen3 235B-A22B | 235.0B (22.0B active) | — | — | — | — | — |
| MiniMax M2.5 | 230.0B (10.0B active) | — | — | — | — | — |
| MiniMax M2.7 | 230.0B (10.0B active) | — | — | — | — | — |
| Mixtral 8x22B | 141.0B (39.0B active) | — | — | — | — | — |
| DBRX 132B | 132.0B (36.0B active) | — | — | — | — | — |
| Nemotron-3 Super 120B-A12B | 120.0B (12.0B active) | — | — | — | — | — |
| Mistral Small 4 119B | 119.0B (11.0B active) | — | — | — | — | — |
| GPT OSS 120B | 117.0B (5.1B active) | — | — | — | — | — |
| Llama 4 Scout (109B) | 109.0B (17.0B active) | — | — | — | — | — |
| Qwen3-Next 80B-A3B | 80.0B (3.0B active) | — | — | — | — | — |
| Falcon 40B | 40.0B | — | — | — | — | — |
| Falcon 7B | 7.0B | — | — | — | — | — |
Frequently asked questions
What is the largest LLM that fits on a single RTX A4000 16GB?
At 8K context and batch 1, the largest models in our catalog that fit on an RTX A4000 16GB (16 GB) are: Qwen3-VL 4B (4.44B) at BF16; GLM-4 9B (9B) at FP8; DeepSeek Coder V2 Lite (16B) at 4-bit (AWQ). Larger models need tensor parallelism across several GPUs or lower-precision weights.
Can an RTX A4000 16GB run LLaMA 3.3 70B?
at BF16: does not fit within one 8-GPU node; at FP8: needs 8× RTX A4000 16GB; at 4-bit (AWQ): needs 4× RTX A4000 16GB. 8K context, batch 1, vLLM — the calculator sizes other contexts and concurrency.
Can an RTX A4000 16GB serve DeepSeek R1?
at BF16: does not fit within one 8-GPU node; at FP8: does not fit within one 8-GPU node; at 4-bit (AWQ): does not fit within one 8-GPU node. 8K context, batch 1, vLLM — the calculator sizes other contexts and concurrency.
How fast is an RTX A4000 16GB for LLM inference?
Single-stream decode for LLaMA 3.1 8B at FP8 is about 30 tokens/s on one RTX A4000 16GB; decode is bound by memory bandwidth (448 GB/s), so larger models run proportionally slower. Batching raises total throughput far above the single-stream figure.
How much does it cost to rent an RTX A4000 16GB?
For on-demand rental our catalog estimate is $0.08/hr. The $/1M tokens column above uses that price.
Methodology: every cell comes from the same engine as our GPU Lookup — weights, KV cache (including MLA latent and hybrid / sliding-window layers), activations and framework overhead per model architecture, decode speed from a calibrated memory-bandwidth roofline. Reference scenario: 8K context (4K for models with a shorter native window), batch 1, vLLM, fp16 KV cache. Fit bands: optimal 50–70% of VRAM, good 70–80%, tight 80–95%. Engineering estimates, not a guarantee — validate against your serving framework before buying hardware.