96 GB VRAM 4000 GB/s Bandwidth
148 TFLOPS FP16 · 400W
NVIDIA H20 96GB (HBM3) Local LLM Benchmark & Guide
Runs 18 open-weight models cleanly · Ref $14,000 (~$146/GB VRAM)
Top Runnable Models on H20 96G by Use Case
Click to load ready-to-run CLI config Best for Coding & Agents
~359.4 tok/s
Qwen3.5-122B-A10B (MoE) Coding score 84 · q5_k_m (90.33/94.1 GB)
Best for General Chat & Reasoning
~359.4 tok/s
Qwen3.5-122B-A10B (MoE) Arena Elo ~1478 · q5_k_m (90.33/94.1 GB)
Best for Vision & OCR
~359.4 tok/s
Qwen3.5-122B-A10B (MoE) Supports Text+Image+Video · q5_k_m (90.33/94.1 GB)
Concurrency (Batch):
MoE 122B / A10B·48L (d=3072)·32Q/2KV·Ctx 256K·Text+Image+Video
Smooth Fit77.26 / 94.1 GB~420 tok/s· Q4_K_M · 16K ctx
3. Weight Quantization
4. Per-Request Context16K (16,384 tok)
Verified: 2026-11-20·Roofline bandwidth/compute estimate (Batch=1)
Smooth Fit
77.26/ 94.1 GB usable
Physical 96 GB · Load 82.1%
Theoretical Decode Speed
~420tokens/s
MoE: reads only 6.06 GB active weights/step
Max Context at Batch=1
256K/ 256K
Enable Q8/Q4 KV quantization to scale batch concurrency
Weights: 73.96 GBKV Cache (1×16K): 0.52 GBOverhead: 2.78 GB
🚀 Blazing Fast (Instant CoT)~1.2s for a 500-token response
Reference Launch CLI:
OLLAMA_CONTEXT_LENGTH=16384 ollama run qwen3.5:122b-a10bH20 96G Fits Qwen3.5-122B-A10B (MoE) · Cost & TCO
Approx. $14,000 (~$146/GB VRAM) · 400W TDP
Checking current market prices for NVIDIA H20 96GB (HBM3)? Search Price on Amazon →
📋 All Open-Weight Models Tested on H20 96G (Q4_K_M, 16K)
| Model | Q4 VRAM | Fit Status | Est. Speed | Best Quant | Report |
|---|---|---|---|---|---|
| Qwen3.5-122B-A10B (MoE) | 77.26 GB | Smooth Fit | ~420 tok/s | q5_k_m | View → |
| Qwen3.5-35B-A3B (MoE) | 23.34 GB | Smooth Fit | ~420 tok/s | fp16 | View → |
| Qwen3.5-27B | 19.33 GB | Smooth Fit | ~169 tok/s | fp16 | View → |
| Gemma-4-31B-IT | 22.47 GB | Smooth Fit | ~146.2 tok/s | fp16 | View → |
| Gemma-4-26B-A4B-IT (MoE) | 17.92 GB | Smooth Fit | ~420 tok/s | fp16 | View → |
| Llama-4-Scout-109B-A17B (MoE) | 71.92 GB | Smooth Fit | ~241.7 tok/s | q5_k_m | View → |
| Qwen3-32B | 25.03 GB | Smooth Fit | ~130.7 tok/s | fp16 | View → |
| DeepSeek-R1-Distill-Qwen-32B | 25.03 GB | Smooth Fit | ~130.7 tok/s | fp16 | View → |
| GLM-4-32B-0414 | 21.26 GB | Smooth Fit | ~144.6 tok/s | fp16 | View → |
| Qwen3-30B-A3B (MoE) | 20.92 GB | Smooth Fit | ~420 tok/s | fp16 | View → |
| Llama-3.3-70B-Instruct | 49.48 GB | Smooth Fit | ~63.3 tok/s | q8_0 | View → |
| Qwen3-14B | 12.29 GB | Smooth Fit | ~279.2 tok/s | fp16 | View → |
| DeepSeek-R1-Distill-Qwen-14B | 12.82 GB | Smooth Fit | ~272.1 tok/s | fp16 | View → |
| Qwen3-8B | 7.94 GB | Smooth Fit | ~420 tok/s | fp16 | View → |
| Gemma-4-E4B-IT (Edge Multimodal) | 6.25 GB | Smooth Fit | ~420 tok/s | fp16 | View → |
| DeepSeek-R1-Distill-Qwen-7B | 6.09 GB | Smooth Fit | ~420 tok/s | fp16 | View → |
| Qwen3.6-35B-A3B (MoE) | 23.96 GB | Smooth Fit | ~420 tok/s | fp16 | View → |
| Qwen3.8-27B | 19.83 GB | Smooth Fit | ~164.4 tok/s | fp16 | View → |
| Step-3.5-Flash (196B MoE) | 123.02 GB | Partial RAM Offload (Slow) | ~37.8 tok/s | q2_k | View → |
| Qwen3.8-Flash-Next (180B MoE) | 113.11 GB | Partial RAM Offload (Slow) | ~71 tok/s | q3_k_m | View → |
| Kimi K3 (2.8T MoE) | 1733.2 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| DeepSeek-V4-Pro (1.6T MoE) | 990.02 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| GLM-5 (744B-A40B MoE) | 461.99 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| Kimi K2.6 (1T MoE) | 620.83 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| Qwen3.5-397B-A17B (MoE) | 247.5 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| DeepSeek-R1 (671B-A37B MoE) | 416.53 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| GLM-5.3-Flash (320B-A18B MoE) | 199.39 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| Qwen3-235B-A22B (MoE) | 148.98 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| DeepSeek-V4-Flash (284B-A13B MoE) | 176.19 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| MiniMax-M3.1-Flash (428B MoE) | 267.98 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| Kimi K3.1 Pro (2.8T MoE) | 1722.14 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| Qwen3.8-2.4T-A95B (MoE) | 1515.15 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| Meituan LongCat-2.5 (1.78T MoE) | 1101.87 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| Meituan LongCat-2.0 (1.78T MoE) | 1099.4 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| Kimi K2.7 Code (1T MoE) | 636.9 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| Kimi K2.5 (1T MoE) | 636.9 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| Xiaomi MiMo-V2.6-Pro (1T MoE) | 641.56 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| Mistral Large 4 (1T-A52B) | 621.03 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| GLM-5.3 (753B MoE) | 467.56 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| GLM-5.2 (753B MoE) | 467.56 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| Step 5 Preview (600B MoE) | 376.18 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| MiniMax-M3 (427B MoE) | 267.36 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| DeepSeek-V4.1-Flash (1017) | 236.39 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| Xiaomi MiMo-V2.6-Flash (311B MoE) | 196.03 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| ERNIE-5.5-Flash (280B MoE) | 178.12 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| MiniMax-M2.7 (229B MoE) | 146.22 GB | Out of VRAM (OOM) | — | Cloud / Multi-GPU | View → |
| Step 3.7 Flash (201B MoE) | 126.47 GB | Out of VRAM (OOM) | — | q2_k | View → |
Frequently Asked Questions
What is the largest LLM NVIDIA H20 96GB (HBM3) can run smoothly?
With 96 GB VRAM, H20 96G comfortably runs 4-bit (Q4_K_M) models up to Qwen3.5-122B-A10B (MoE) (122B parameters).
How fast is H20 96G for LLM token generation?
Driven by its 4000 GB/s memory bandwidth, H20 96G generates roughly 420 tokens/s on Qwen3.5-122B-A10B (MoE) (Q4_K_M).