Can the RTX 3090 Run DeepSeek-R1-Distill-Qwen 14B?
Runs
Yes. DeepSeek-R1-Distill-Qwen 14B at Q4_K_M is a 9 GB download (Hugging Face file listing for bartowski/DeepSeek-R1-Distill-Qwen-14B-GGUF), and the RTX 3090 has 24 GB of VRAM, leaving about 15 GB for the KV cache and runtime. One published run puts generation at 37.5 tokens per second; a single cited measurement, not a median. Cited run: SmeltCore.
Buy the RTX 3090
ZOTAC Gaming GeForce RTX 3090 Trinity OC 24GB GDDR6X 384-bit 19.5 Gbps PCIE 4.0 Gaming…Tracked at $1,900 on 2026-11-07 — price may vary.
Published runs for DeepSeek-R1-Distill-Qwen 14B on the RTX 3090
| Generation | Quant | Runtime | Context | Source | Date |
|---|---|---|---|---|---|
| 37.5 tok/s | q4_K_M | ollama | 16K | SmeltCorelabelled "deepseek-r1:14b" | 2026-05-09 |
More about this pair
How much VRAM does DeepSeek-R1-Distill-Qwen 14B need?
The Q4_K_M file is 9 GB. Add roughly 1.5 GB for the runtime and an ~8K-token context, so about 10.5 GB of VRAM holds it comfortably. Longer contexts grow the KV cache and need more.
What speed does DeepSeek-R1-Distill-Qwen 14B reach on the RTX 3090?
The fastest published run is 37.5 tokens per second (SmeltCore, q4_K_M, ollama). Generation reads every weight once per token, so the RTX 3090's 936 GB/s of memory bandwidth divided by the 8.99 GB file puts the theoretical ceiling near 104 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
DeepSeek-R1-Distill-Qwen 14B on other GPUs
- RTX 4060 — Offload
- RTX 3060 Ti — Offload
- RTX 3060 12GB — Runs, 29.4 tok/s (1 run)
- Intel Arc B580 — Runs
- RTX 4070 SUPER — Runs
- RTX 4060 Ti 16GB — Runs
- RTX 5060 Ti 16GB — Runs
- Intel Arc A770 16GB — Runs, 13.4 tok/s (1 run)
- Radeon RX 9070 XT — Runs
- RTX 4070 Ti SUPER — Runs
- RTX 4080 — Runs
- RTX 5070 Ti — Runs
- RTX 5080 — Runs, 70 tok/s (1 run)
- Tesla P40 24GB — Runs
- Radeon RX 7900 XTX — Runs
- RTX 4090 — Runs, 59 tok/s median
- RTX 5090 — Runs, 89.1 tok/s (1 run)
- RTX A6000 48GB — Runs, 48.4 tok/s (1 run)
- RTX PRO 5000 Blackwell 48GB — Runs, 110 tok/s (1 run)
- Instinct MI210 64GB — Runs
- RTX PRO 6000 Blackwell 96GB — Runs
Other models on the RTX 3090
- Llama 3.2 1B — Runs, 267 tok/s (1 run)
- Llama 3.2 3B — Runs
- Mistral 7B — Runs
- Qwen2.5 7B — Runs
- Llama 3.1 8B — Runs, 92 tok/s median
- DeepSeek-R1-Distill-Llama 8B — Runs
- Qwen3 8B — Runs, 101 tok/s median
- Gemma 2 9B — Runs
- Gemma 3 12B — Runs
- Qwen2.5 14B — Runs, 55.8 tok/s median
- Qwen3 14B — Runs, 61.1 tok/s median
- Phi-4 14B — Runs
- gpt-oss-20b — Runs
- Gemma 3 27B — Runs
- Gemma 2 27B — Runs
- Qwen3 30B-A3B — Runs
- Qwen3 32B — Runs, 32.7 tok/s median
- Qwen2.5 32B — Runs, 23 tok/s median
- DeepSeek-R1-Distill-Qwen 32B — Runs
- Qwen3 32B (Q6_K) — Offload
- Qwen2.5 32B (Q6_K) — Offload
- DeepSeek-R1-Distill-Qwen 32B (Q6_K) — Offload
- Gemma 3 27B (Q8_0) — Offload
- Llama 3.1 70B — Offload
- Llama 3.3 70B — Offload
- DeepSeek-R1-Distill-Llama 70B — Offload
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.