ai.canshu.com
中文
VRAM Calculator Advisor & TCO API Pricing Compare Models Leaderboard GPUs

Search models, GPUs, and tools

↑↓ Navigate · Enter Open · ★ Watchlist prioritized ai.canshu.com
Q4_K_M: Smooth Fit 12.82 GB / 94.1 GB Usable

DeepSeek-R1-Distill-Qwen-14B on NVIDIA H20 96GB (HBM3): VRAM Requirement & Speed Estimate

Yes. At Q4_K_M 4-bit quantization with a 16K context window, DeepSeek-R1-Distill-Qwen-14B requires 12.82 GB of VRAM (8.97 GB weights + 3.22 GB KV cache + 0.63 GB overhead), fitting within H20 96G's 94.1 GB usable memory at ~272.1 tokens/s.

Concurrency (Batch):
Dense 14.8B·48L (d=5120)·40Q/8KV·Ctx 128K
Runnable Models →
GPUs
96 GB VRAM·4000 GB/s·Safe Usable 94.1 GB
Smooth Fit12.82 / 94.1 GB~272.1 tok/s
3. Weight Quantization
4. Per-Request Context16K (16,384 tok)
Verified: 2026-11-20·Roofline bandwidth/compute estimate (Batch=1)
Smooth Fit
12.82/ 94.1 GB usable

Physical 96 GB · Load 13.6%

Theoretical Decode Speed
~272.1tokens/s

Actual speed varies ~±25% by runtime

Max Context at Batch=1
128K/ 128K

Enable Q8/Q4 KV quantization to scale batch concurrency

Weights: 8.97 GBKV Cache (1×16K): 3.22 GBOverhead: 0.63 GB
🚀 Blazing Fast (Instant CoT)~1.8s for a 500-token response
Reference Launch CLI:
OLLAMA_CONTEXT_LENGTH=16384 ollama run deepseek-r1:14b

H20 96G Fits DeepSeek-R1-Distill-Qwen-14B · Cost & TCO

Approx. $14,000 (~$146/GB VRAM) · 400W TDP

📊 All Quantization Levels for DeepSeek-R1-Distill-Qwen-14B on H20 96G

Quantization Total VRAM (16K) Verdict Est. Speed Max Context
FP16 / BF16 33.86 GB Smooth Fit ~92.3 tok/s 128K
FP8 / Q8_0 19.71 GB Smooth Fit ~166.1 tok/s 128K
GGUF Q6_K 16.05 GB Smooth Fit ~209.5 tok/s 128K
GGUF Q5_K_M 14.41 GB Smooth Fit ~237.3 tok/s 128K
Q4_K_M / AWQ 12.82 GB Smooth Fit ~272.1 tok/s 128K
GGUF IQ4_XS 11.69 GB Smooth Fit ~304 tok/s 128K
GGUF Q3_K_M 11.05 GB Smooth Fit ~325.6 tok/s 128K
GGUF Q2_K 9.62 GB Smooth Fit ~387.2 tok/s 128K

💡 Recommended Alternative GPUs & Macs for DeepSeek-R1-Distill-Qwen-14B

Frequently Asked Questions

Can NVIDIA H20 96GB (HBM3) run DeepSeek-R1-Distill-Qwen-14B locally?

Yes. At Q4_K_M 4-bit quantization with a 16K context window, DeepSeek-R1-Distill-Qwen-14B requires 12.82 GB of VRAM (8.97 GB weights + 3.22 GB KV cache + 0.63 GB overhead), fitting within H20 96G's 94.1 GB usable memory at ~272.1 tokens/s.

What is the best quantization for DeepSeek-R1-Distill-Qwen-14B on H20 96G?

The highest-quality quantization that fits cleanly on H20 96G is FP16 / BF16 (Lossless Full Precision), using 33.86 GB VRAM at ~92.3 tokens/s with up to 128K context.

How much VRAM does DeepSeek-R1-Distill-Qwen-14B need in full FP16 precision?

In unquantized FP16/BF16 precision (16K context), DeepSeek-R1-Distill-Qwen-14B requires 33.86 GB VRAM (29.6 GB for weights alone).