Can the RTX A6000 48GB Run Phi-4 14B?
Runs
Yes. Phi-4 14B at Q4_K_M is a 9.1 GB download (Hugging Face file listing for bartowski/phi-4-GGUF), and the RTX A6000 48GB has 48 GB of VRAM, leaving about 39 GB for the KV cache and runtime. One published run puts generation at 52.6 tokens per second; a single cited measurement, not a median. Cited run: DatabaseMart.
Buy the RTX A6000 48GB
No Amazon listing for this card is tracked today — see live eBay listings.
Published runs for Phi-4 14B on the RTX A6000 48GB
| Generation | Quant | Runtime | Context | Source | Date |
|---|---|---|---|---|---|
| 52.6 tok/s | q4_K_M | ollama | — | DatabaseMartlabelled "phi4:14b" | 2026-07-08 |
More about this pair
How much VRAM does Phi-4 14B need?
The Q4_K_M file is 9.1 GB. Add roughly 1.5 GB for the runtime and an ~8K-token context, so about 10.6 GB of VRAM holds it comfortably. Longer contexts grow the KV cache and need more.
What speed does Phi-4 14B reach on the RTX A6000 48GB?
The fastest published run is 52.6 tokens per second (DatabaseMart, q4_K_M, ollama). Generation reads every weight once per token, so the RTX A6000 48GB's 768 GB/s of memory bandwidth divided by the 9.05 GB file puts the theoretical ceiling near 85 tokens per second. That is an upper bound, not a speed estimate: kernel, launch and compute overhead keep real runs below it, and SpecPicks rejects any published figure above it.
Check another pair
Phi-4 14B on other GPUs
- RTX 4060 — Offload
- RTX 3060 Ti — Offload
- RTX 3060 12GB — Runs
- Intel Arc B580 — Runs
- RTX 4070 SUPER — Runs
- RTX 4060 Ti 16GB — Runs
- RTX 5060 Ti 16GB — Runs
- Intel Arc A770 16GB — Runs, 22.3 tok/s median
- Radeon RX 9070 XT — Runs
- RTX 4070 Ti SUPER — Runs
- RTX 4080 — Runs
- RTX 5070 Ti — Runs
- RTX 5080 — Runs
- Tesla P40 24GB — Runs
- RTX 3090 — Runs
- Radeon RX 7900 XTX — Runs
- RTX 4090 — Runs, 68.8 tok/s median
- RTX 5090 — Runs, 130 tok/s (1 run)
- RTX PRO 5000 Blackwell 48GB — Runs
- Instinct MI210 64GB — Runs
- RTX PRO 6000 Blackwell 96GB — Runs
Other models on the RTX A6000 48GB
- Llama 3.2 1B — Runs
- Llama 3.2 3B — Runs
- Mistral 7B — Runs
- Qwen2.5 7B — Runs
- Llama 3.1 8B — Runs
- DeepSeek-R1-Distill-Llama 8B — Runs
- Qwen3 8B — Runs
- Gemma 2 9B — Runs
- Gemma 3 12B — Runs
- Qwen2.5 14B — Runs, 50.3 tok/s (1 run)
- DeepSeek-R1-Distill-Qwen 14B — Runs, 48.4 tok/s (1 run)
- Qwen3 14B — Runs, 40.7 tok/s (1 run)
- gpt-oss-20b — Runs
- Gemma 3 27B — Runs
- Gemma 2 27B — Runs, 31.6 tok/s (1 run)
- Qwen3 30B-A3B — Runs
- Qwen3 32B — Runs, 27.5 tok/s median
- Qwen2.5 32B — Runs, 26.1 tok/s (1 run)
- DeepSeek-R1-Distill-Qwen 32B — Runs, 26.2 tok/s (1 run)
- Qwen3 32B (Q6_K) — Runs
- Qwen2.5 32B (Q6_K) — Runs
- DeepSeek-R1-Distill-Qwen 32B (Q6_K) — Runs
- Gemma 3 27B (Q8_0) — Runs
- Llama 3.1 70B — Runs, 14.6 tok/s (1 run)
- Llama 3.3 70B — Runs, 13.6 tok/s (1 run)
- DeepSeek-R1-Distill-Llama 70B — Runs, 13.9 tok/s median
How the checks work
The fit is the Q4_K_M file size against the card's VRAM, with 1.5 GB kept for the runtime and an ~8K context. A published run counts only if the pair fits, the run is Q4-class, its source does not call it an estimate, it is under the bandwidth ceiling (memory bandwidth ÷ GB read per token: the whole file for a dense model, only the active experts' weights for a mixture-of-experts model such as gpt-oss-20b or Qwen3 30B-A3B, which is why those can post figures above bandwidth ÷ file size), it is not slower than the same card on a model twice its size, a single run is not slower than the same card on a larger dense model, and it is not one figure pasted across many cards. The whole matrix is downloadable as a CSV under CC BY 4.0.
As an Amazon Associate, SpecPicks earns from qualifying purchases. As an eBay Partner Network member, SpecPicks may earn a commission on eBay links.
This page is editorial synthesis based on publicly available information. No independent first-party benchmarking is reported.