DeepSeek V3.2 (685B) on Gaudi 3 128GB HBM2e (128GB) — VRAM, KV Cache & Throughput

DeepSeek V3.2 (685B) (685.36B total, 37B active, Multi-head Latent Attention (MLA) MoE) uses 377.56 GiB of 4-bit weights (Q4_K_M), requiring 378.79 GiB total at 8K context and 380.75 GiB at 32K on Gaudi 3 128GB HBM2e (~3.6 tok/s decode).

Interactive Sizing Workbench

Full Compatibility Report →

Open-Weight Model685.36B / 37B

Weights & QuantizationQ4_K_M · 377.56 GiB

Context Window8,192 (8K)

Target Hardware

128 GB · 3700 GB/s

GPU Count (Tensor Parallel)

Concurrent Batch Size

KV Cache Precision

Inference Engine

System RAM Offload (CPU)0%

Exceeds Available GPU Memory

Measured GGUF shard size HF config.json ↗

Memory Allocation Breakdown (GiB) 378.79 GiB (406.72 GB) / 128 GiB (296%)

Model Weights: 377.56 GiB KV Cache & State: 0.66 GiB Runtime & Activation Buffers: 0.57 GiB System RAM Offload: 0 GiB

Required GPU Memory

378.79 GiB

406.72 GB (decimal)

Usable Hardware Memory

121.6 GiB

128 GB (physical)

Safe Headroom

-257.19 GiB

296% utilization

Est. Decode Throughput

~3.6 tok/s

37B active params

Est. Prefill Throughput

~12895 tok/s

4K Prompt TTFT: ~318 ms

Est. Cost / 1M Tokens

$272.84

Multi-head Latent Attention (MLA) MoE · MIT

Recommended Hardware for Current Configuration

Best Single GPU

Multi-GPU Required

—

Best Value / Used GPU

4x H200 141GB SXM

564GB · ~$128,000

Apple Unified Memory

Mac Studio M3 Ultra 512GB

~$9,500

Cloud From

Vast.ai · $3.54 / hr

8x A100 80GB SXM

Production Launch Commands

llama.cpp (GGUF)

llama-server -hf unsloth/DeepSeek-V3.2-GGUF:Q4_K_M \
  --ctx-size 8192 -ngl 999 --flash-attn

vLLM

vllm serve deepseek-ai/DeepSeek-V3.2 \
  --max-model-len 8192 --gpu-memory-utilization 0.92

Live Token Speed Simulator (TTFT ~318ms · ~3.6 tok/s)

[DeepSeek V3.2 (685B) · 378.79 GiB] allocated on target GPU. Prefill completed in TTFT window; streaming autoregressive tokens at memory-bandwidth-bound rate...

Cloud GPU Comparison & Rent-vs-Buy TCO

Verified public rates across RunPod, Vast.ai, Vultr, and DigitalOcean.

Verified: 2026-11-05

Active usage (hours / day)4 h / day

Local hardware budget ($)

Electricity rate ($/kWh)

Purchase break-even

~9.3 months

Est. local power: $12.77 / mo

Provider	Recommended Instance	Total VRAM	Hourly Rate	Monthly at Selected Hours	Billing Mode	Deploy
RunPod	2x B300 288GB Blackwell Ultra (Community Cloud)	576 GB	$13.88 / hr	$1687.81 / mo	Per-second	Deploy GPU ↗
Vast.ai	8x A100 80GB SXM (Verified Host · SXM4)	640 GB	$3.54 / hr	$429.98 / mo	Spot / On-demand	Deploy GPU ↗
*Vultr ($300 Free Credit)**	8x A100 80GB SXM (Dedicated Instance · $300 Credit*)	640 GB	$19.18 / hr	$2331.80 / mo	Hourly	Deploy GPU ↗
Vultr Cloud GPU	8x A100 80GB SXM (Dedicated Instance)	640 GB	$19.18 / hr	$2331.80 / mo	Hourly	Deploy GPU ↗
DigitalOcean	8x Instinct MI300X 192GB (Dedicated Instance)	1536 GB	$20.72 / hr	$2519.55 / mo	Per-second	Deploy GPU ↗

Rates reflect public listings on the verification date and fluctuate with regional inventory.

Fine-Tuning VRAM Requirements (QLoRA / LoRA / Full) — DeepSeek V3.2 (685B)

QLoRA (4b) Exceeds Available GPU Memory

383.91 GiB

Model Weights: 377.56 GiB · Gradients + Optimizer: 5.18 GiB · Act: 1.17 GiB

Recommended GPU: 4x H200 141GB SXM (564GB)

LoRA (16b) Exceeds Available GPU Memory

1259.8 GiB

Model Weights: 1250.09 GiB · Gradients + Optimizer: 8.14 GiB · Act: 1.57 GiB

Recommended GPU: 8x B200 180GB (1440GB)

Full BF16 Exceeds Available GPU Memory

5128.14 GiB

Model Weights: 1250.09 GiB · Gradients + Optimizer: 3875.28 GiB · Act: 2.77 GiB

Recommended GPU: Multi-GPU Required

Engineering Sizing Notes & Optimization Playbook

Quick Hardware Verdict

A single Gaudi 3 128GB HBM2e (128GB) cannot hold 4-bit DeepSeek V3.2 (685B) (378.79 GiB at 8K) entirely in VRAM. Use multi-GPU tensor parallelism, offload layers to system RAM, or deploy on a larger cloud GPU.

KV Cache & Runtime Tuning

DeepSeek V3.2 (685B) uses Multi-head Latent Attention (MLA) MoE (685.36B total, 37B active per token). Moving from 8K to 32K context expands FP16 KV cache from 0.66 GiB to 2.62 GiB; enabling 8-bit KV cache (llama.cpp -ctk q8_0 -ctv q8_0 or vLLM --kv-cache-dtype fp8) saves 1.31 GiB at 32K (reducing total VRAM to 379.44 GiB).

VRAM Scaling Across Context Windows (4K to 128K, 4-bit Weights)

Model weights remain constant while attention KV cache grows with context length and batch size.

Quantization Precision & Memory Matrix — DeepSeek V3.2 (685B) @ 8K

Quantization Format	Weights (GiB)	KV Cache	Total VRAM (GiB)	Decode Speed	Compatibility (Gaudi 3 128GB HBM2e)	Recommended GPU	Data Source
Native Weights (BF16 / FP8 / MXFP4 — Lossless) (FP8)	1250.09 GiB	0.66 GiB	1251.32 GiB (1343.59 GB)	~0.8 tok/s	Exceeds Available GPU Memory	Multi-GPU Required	Official checkpoint size
FP8 / GGUF Q8_0 (8-bit — Near-Lossless) (Q8_0)	664.3 GiB	0.66 GiB	665.53 GiB (714.61 GB)	~1.7 tok/s	Exceeds Available GPU Memory	Multi-GPU Required	Measured GGUF shard size
GGUF Q6_K (6-bit — High Precision) (Q6_K)	513.45 GiB	0.66 GiB	514.68 GiB (552.63 GB)	~2.4 tok/s	Exceeds Available GPU Memory	Multi-GPU Required	Measured GGUF shard size
GGUF Q5_K_M (5-bit — Balanced) (Q5_K_M)	443.46 GiB	0.66 GiB	444.69 GiB (477.48 GB)	~2.9 tok/s	Exceeds Available GPU Memory	Multi-GPU Required	Measured GGUF shard size
Q4_K_M / UD-Q4_K_XL / AWQ (4-bit — Recommended) (Q4_K_M)	377.56 GiB	0.66 GiB	378.79 GiB (406.72 GB)	~3.6 tok/s	Exceeds Available GPU Memory	Multi-GPU Required	Measured GGUF shard size
GGUF IQ4_XS (4.25-bit — Compact) (IQ4_XS)	333.74 GiB	0.66 GiB	334.97 GiB (359.67 GB)	~4.4 tok/s	Exceeds Available GPU Memory	Multi-GPU Required	Measured GGUF shard size
GGUF Q3_K_M (3-bit — Low VRAM) (Q3_K_M)	298.24 GiB	0.66 GiB	299.47 GiB (321.55 GB)	~5.2 tok/s	Exceeds Available GPU Memory	Multi-GPU Required	Measured GGUF shard size
GGUF UD-Q2_K_XL (2-bit — Extreme MoE) (UD-Q2_K_XL)	229.67 GiB	0.66 GiB	230.9 GiB (247.93 GB)	~8.4 tok/s	Exceeds Available GPU Memory	Instinct MI325X 256GB (256GB)	Measured GGUF shard size

Technical Specifications & FAQ

Does DeepSeek V3.2 (685B) fit on Gaudi 3 128GB HBM2e (128GB)?

In 4-bit (Q4_K_M) at 8,192 context, DeepSeek V3.2 (685B) requires 378.79 GiB total GPU memory (377.56 GiB weights + 0.66 GiB KV cache + 0.57 GiB runtime buffers) against 121.6 GiB usable on Gaudi 3 128GB HBM2e, delivering ~3.6 tok/s decode.

How much VRAM does DeepSeek V3.2 (685B) need at 32K context?

At 32,768 context tokens, DeepSeek V3.2 (685B) (Multi-head Latent Attention (MLA) MoE) allocates 2.62 GiB for KV cache and state, bringing total 4-bit VRAM to 380.75 GiB (408.83 GB decimal).

Evaluate This Model on Other GPUs

RTX 3060 12GB (12GB)RTX 4060 Ti 16GB (16GB)RTX 5060 Ti 16GB (16GB)RTX 4070 Ti Super (16GB)RTX 4080 Super (16GB)RTX 5070 12GB (12GB)RTX 5070 Ti (16GB)RTX 5080 (16GB)RTX 3090 (24GB)RTX 4090 (24GB)RTX 4090 D (24GB)RTX 5090 (32GB)RTX 5090 D (32GB)RTX A6000 (48GB)RTX 6000 Ada (48GB)RTX PRO 6000 Blackwell (96GB)DGX Spark (GB10 128GB) (128GB)L4 24GB (24GB)A40 48GB (48GB)L20 48GB (48GB)L40S 48GB (48GB)A800 80GB SXM (80GB)A100 80GB SXM (80GB)H800 80GB SXM (80GB)H100 80GB SXM (80GB)H20 96GB HBM3 (96GB)H20 141GB HBM3e (141GB)H200 141GB SXM (141GB)B200 180GB (180GB)B300 288GB Blackwell Ultra (288GB)Ascend 910B 64GB (昇腾) (64GB)Ascend 910C 128GB (昇腾) (128GB)Atlas 300I Duo 96GB (310P) (96GB)MLU590 80GB (寒武纪思元) (80GB)DCU K100-AI 64GB (海光深算) (64GB)MetaX C500 64GB (沐曦曦云) (64GB)Kunlunxin P800 64GB (昆仑芯) (64GB)MTT S4000 48GB (摩尔线程) (48GB)Radeon RX 7900 XTX 24GB (24GB)Radeon AI PRO R9700 32GB (32GB)Ryzen AI Max+ 395 128GB (128GB)Instinct MI300X 192GB (192GB)Instinct MI325X 256GB (256GB)Instinct MI355X 288GB (288GB)Arc Pro B60 24GB (24GB)Mac mini M4 Pro 64GB (64GB)MacBook Pro M4 Max 128GB (128GB)Mac Studio M4 Ultra 256GB (256GB)Mac Studio M3 Ultra 512GB (512GB)

Evaluate Other Models on This Hardware

Qwen3.8 Flash Next (125B-A6B)Qwen3.8 27B Qwen3.8 2.4T-A95B Qwen3.6 35B-A3B Qwen3.5 397B-A17B Qwen3.5 122B-A10B Qwen3.5 35B-A3B Qwen3.5 27B Qwen3.5 9B Qwen3.5 4B Qwen3 Coder 480B-A35B Qwen3 Coder Next 80B-A3B Qwen3 Coder 30B-A3B Qwen3 30B-A3B Instruct 2507 Qwen3 235B-A22B Instruct 2507 Qwen3 32B Qwen3 14B Qwen3 8B Qwen3 4B Instruct 2507 Qwen3 1.7B Qwen3 0.6B Qwen2.5 72B Instruct Qwen2.5 Coder 32B Instruct QwQ 32B Reasoning DeepSeek V4.1 Flash (552B-A16B)DeepSeek V4 Flash DeepSeek R1-0528 671B DeepSeek V3-0324 671B DeepSeek R1 Distill Llama 70B DeepSeek R1 Distill Qwen 32B DeepSeek R1 Distill Qwen 14B DeepSeek R1 Distill Llama 8B gpt-oss-20b gpt-oss-120b Gemma 4 31B IT Gemma 4 26B-A4B IT Gemma 4 12B IT Gemma 4 E4B IT Gemma 4 E2B IT Gemma 3 27B IT Gemma 3 12B IT Gemma 3 4B IT GLM-5.3 Flash (320B-A18B)GLM-5.3 (753B)GLM-5 (744B)GLM-4.5 (355B-A32B)GLM-4.5 Air (110B-A12B)GLM-4 32B-0414 Kimi K3.1 Pro (2.8T-A104B)Kimi K3 (2.8T-A104B)Kimi K2 Instruct (1T-A32B)Kimi-Dev 72B MiniMax M3.1 Flash (428B-A23B)MiniMax M2.5 (229B-A10B)Step-5 Preview (600B-A27B)Step-3.7 Flash (198B-A11B)Step-3.5 Flash (197B-A11B)ERNIE 4.5 300B-A47B (文心)ERNIE 4.5 21B-A3B (文心)Hunyuan A13B Instruct (混元 80B)MiniCPM5 2B (面壁)Llama 4 Maverick 17B-128E (400B)Llama 4 Scout 17B-16E (109B)Llama 3.1 405B Instruct Llama 3.3 70B Instruct Llama 3.1 8B Instruct Llama 3.2 3B Instruct Mistral Large 4 (1T-A52B)Mistral Large 2 123B Mixtral 8x22B Instruct Mixtral 8x7B Instruct Mistral Small 3.2 24B Devstral Small 24B Nemotron 3 Ultra 550B-A55B Nemotron 3 Super 120B-A12B Nemotron 3 Nano 30B-A3B Cohere Command A 111B Phi-4 14B Liquid LFM3 1.7B