NVIDIA
Operating mode
Use this to bias workload recommendations toward responsiveness, background autonomy, lighter serving, or multi-GPU scale-out.
Current mode
Balanced
Balanced for general local use. Keeps the ranking neutral across personal and serving workflows.
NVIDIA H200 SXM 是升级版 Hopper 旗舰，拥有 141 GB HBM3e 内存——几乎是 H100 容量的两倍——是首款能以 FP16 精度运行 70B 模型并留有大量 KV 缓存空间的单 GPU。其 4800 GB/s 带宽比 H100 SXM 高 43%，可按比例提升内存带宽受限推理的 token 生成速度。计算核心与 H100 相同，因此 H200 的主要优势在于内存容量和带宽而非原始算力。
Beyond LLMs
What AI tasks this GPU can handle — from text generation to image and video creation.
| Capability | Status | Representative Model |
|---|---|---|
| LLM Chat (7B) | Runs natively | Llama 3.1 8B Q4 |
| LLM Coding (30B) | Runs natively | Qwen 3 30B Q4 |
| LLM Large (70B) | Runs natively | Llama 3.1 70B Q4 |
| Image Gen (SDXL) | Runs natively | SDXL 1.0 FP16 |
| Image Gen (Flux) | Runs natively | Flux.1 Dev FP16 |
| Image Gen (SD 3.5) | Runs natively | SD 3.5 Large FP16 |
| Video Short (25f) | Runs natively | LTX Video 2B |
| Video Long (100f) | Runs natively | Wan Video 14B |
Architecture
Hopper is NVIDIA's datacenter-focused architecture succeeding Ampere. Built on TSMC 4N, it introduces the Transformer Engine with automatic FP8/FP16 mixed-precision training, HBM3/HBM3e memory, and NVLink 4.0 for multi-GPU scaling. The H100 flagship delivers up to 3x the AI training performance of A100.
AI Relevance
The Transformer Engine automatically manages FP8 precision for optimal training speed without accuracy loss. With up to 141 GB HBM3e (H200), Hopper GPUs can hold the largest open-weight models entirely in GPU memory, making them the workhorse of AI datacenters.
购买建议
本地 AI 的绝佳选择
能良好运行 50 个顶级模型中的 38 个 — 本地推理的全能之选。
141.0 GB
VRAM
$30,000
建议零售价
$213/GB
每 GB VRAM 成本
最适合此 GPU 的模型
What will limit you first
This setup is broadly balanced for this model.
No major red flags
This recommendation has enough memory headroom and acceptable estimated speed for the selected workload.
Best upgrade itinerary
Unlocks 6 additional models that do not fit on the current setup.
想要更多余量？ NVIDIA B200 180GB (180.0 GB VRAM) 是下一步升级选择。
Cost vs cloud API
Assumes 4 hours/day of active inference at 162 tok/s, NVIDIA H200 141GB amortized over 36 months, US residential electricity ($0.15/kWh), blended cloud pricing at $10 per 1M tokens (GPT-4o / Claude Sonnet tier).
70.0M
Tokens/month at this pace
$985
Monthly local cost
$700
Same tokens on cloud API
$14.1
Local $/1M tokens
Break-even: long amortization at this workload — local is still the privacy/latency play. Price reference: $35.0k MSRP.
Chat
SQwen 3.5 122B A10B matches Chat and keeps a practical fit profile. It is a recent-generation family, which helps on current local SOTA workloads. It fits natively with comfortable headroom. Context coverage stays within the requested workload envelope. Known distribution channels: huggingface, lm-studio.
Coding
SQwen 3.5 122B A10B is a specialized fit for Coding. It is a recent-generation family, which helps on current local SOTA workloads. It fits natively with comfortable headroom. Context coverage stays within the requested workload envelope. Known distribution channels: huggingface, lm-studio.
Agentic Coding
SDevstral 2 123B Instruct is a specialized fit for Agentic Coding. It is a recent-generation family, which helps on current local SOTA workloads. It fits natively with comfortable headroom. Context coverage stays within the requested workload envelope. Known distribution channels: huggingface, lm-studio.
Reasoning
SDevstral 2 123B Instruct matches Reasoning and keeps a practical fit profile. It is a recent-generation family, which helps on current local SOTA workloads. It fits natively with comfortable headroom. Context coverage stays within the requested workload envelope. Known distribution channels: huggingface, lm-studio.
RAG
SQwen 3.5 122B A10B matches RAG and keeps a practical fit profile. It is a recent-generation family, which helps on current local SOTA workloads. It fits natively with comfortable headroom. Context coverage stays within the requested workload envelope. Known distribution channels: huggingface, lm-studio.
Qwen 3.5 122B A10B
Devstral 2 123B InstructMistral Small 4 119B
GPT-OSS 120BPixtral Large 124B
Command A 111BLeanstral 119B A6BQwen 2.5 VL 72BQwen3-Coder-NextQwen3-Coder 30B A3B InstructQwen 3.6 35B A3BQwen 3.5 27BQwen3-VL 30B A3B InstructQwen 3.6 27BQwen 3.5 35B A3BQwen 3 32BMagistral Small 2507Devstral Small 2 24B InstructQwen 3 30B A3BNemotron 3 Nano 30BQwen 3.5 9BQwen 3 14BDevstral Small 1.1
Gemma 4 31B
Phi-4-reasoning-plus 14BQwen 3 8BNemotron Cascade 2 30B A3BGPT-OSS 20BQwen 3.5 4B
EXAONE 4.0 32BGemma 4 26B A4BMinistral 3 14BQwen 3 235B A22BNemotron Nano 8BPhi-4 Mini Reasoning 4B
Jina Embeddings v3
BGE M3Qwen 3.5 397B A17B
Kimi K2.5Kimi K2.6
DeepSeek V4 ProDeepSeek V4 Flash
GLM-5.1GLM-5DeepSeek V3.2Qwen3-Coder 480B A35B InstructDeepSeek Coder V2 236BDeepSeek R1 671BDeepSeek V3.1 671B触手可及
高质量模型，只需稍多一点内存
Image & Video Generation
52 of 52 models can generate images or video on your NVIDIA H200 141GB
| Model | Max Resolution | Gen Time | Grade |
|---|---|---|---|
| SD TurboImage | 512×512 | 0ms | S |
| Stable Diffusion 1.5Image | 512×768 | 100ms | S |
| Realistic Vision v5.1Image | 512×768 | 100ms | S |
| DreamShaper 8Image | 512×768 | 100ms | S |
| LCM DreamShaper v7Image | 512×768 | 0ms | S |
| PixArt-SigmaImage | 1024×1024 | 300ms | S |
| FramePack I2VVideo | 1280×720 | 600ms/frame | S |
| SDXL TurboImage | 512×512 | 0ms | S |
| SDXL LightningImage | 1024×1024 | 100ms | S |
| Stable Diffusion XL 1.0Image | 1024×1024 | 300ms | S |
| Playground v2.5Image | 1024×1024 | 500ms | S |
| RealVisXL v5.0Image | 1024×1024 | 300ms | S |
| DreamShaper XLImage | 1024×1024 | 300ms | S |
| Juggernaut XL v9Image | 1024×1024 | 300ms | S |
| Animagine XL 3.1Image | 1024×1024 | 300ms | S |
| Pony Diffusion V6 XLImage | 1024×1024 | 300ms | S |
| Animagine XL 4.0Image | 1024×1024 | 300ms | S |
| Illustrious XLImage | 1024×1024 | 300ms | S |
| Wan Video 2.1 1.3BVideo | 480×832 | 200ms/frame | S |
| Stable Diffusion 3.5 MediumImage | 1024×1024 | 500ms | S |
| Flux.2 Klein 4BImage | 1024×1024 | 100ms | S |
| LTX Video 2BVideo | 1280×720 | 300ms/frame | S |
| KolorsImage | 1024×1024 | 600ms | S |
| Stable CascadeImage | 1024×1024 | 800ms | S |
| AuraFlow v0.3Image | 1536×1536 | ~1.4s | S |
| Stable Diffusion 3.5 LargeImage | 1024×1024 | ~1.7s | S |
| Stable Diffusion 3.5 Large TurboImage | 1024×1024 | 300ms | S |
| CogVideoX 2BVideo | 720×480 | 300ms/frame | S |
| HunyuanVideoVideo | 720×1280 | 600ms/frame | S |
| ChromaImage | 1024×1024 | 300ms | S |
| Z-Image TurboImage | 1536×1536 | 300ms | S |
| Flux.1 DevImage | 1024×1024 | ~1.4s | S |
| Flux.1 SchnellImage | 1024×1024 | 300ms | S |
| LTX Video 13BVideo | 1280×720 | 600ms/frame | S |
| Flux.1 Kontext DevImage | 1024×1024 | ~1.5s | S |
| AnimateDiff v1.5.3Video | 512×768 | 100ms/frame | S |
| Cosmos Diffusion 7BVideo | 1024×576 | 400ms/frame | S |
| CogVideoX 5BVideo | 720×480 | 400ms/frame | S |
| Wan2.2 TI2V 5BVideo | 832×480 | 400ms/frame | S |
| Flux.2 Klein 9BImage | 1024×1024 | 200ms | S |
| Flux.1 Fill DevImage | 1024×1024 | ~1.3s | S |
| Mochi 1 PreviewVideo | 848×480 | 500ms/frame | S |
| HunyuanVideo 1.5Video | 720×1280 | 500ms/frame | S |
| Helios 14BVideo | 1280×720 | 600ms/frame | S |
| SkyReels V2 14BVideo | 1280×720 | 600ms/frame | S |
| Wan Video 2.1 14BVideo | 720×1280 | 600ms/frame | S |
| Wan Video 2.2 14BVideo | 720×1280 | 600ms/frame | S |
| Qwen ImageImage | 1024×1024 | 500ms | S |
| Qwen Image EditImage | 1024×1024 | 500ms | S |
| Flux.2 DevImage | 1024×1024 | ~14.2s | S |
| MAGI-1Video | 1280×720 | 700ms/frame | S |
| HunyuanImage 3.0Image | 256×256 | 900ms | D |
Image models estimated at 1024×1024 (28 steps, FP16). Video models estimated at 768×512 (25 frames, 30 steps, FP16). Actual performance varies with runtime and system load.
Multi-GPU scaling
Scale out with multiple GPUs for larger models. NVLink provides 900 GB/s inter-GPU bandwidth with 8% overhead.
| Config | Effective memory | Models that fit | Est. bandwidth |
|---|---|---|---|
| 1× NVIDIA | 141 GB | 353/374 | 4,800 GB/s |
| 2× NVIDIA | 282 GB | 364/374 | 8,832 GB/s |
| 4× NVIDIA | 564 GB | 373/374 | 17,664 GB/s |
| 8× NVIDIA | 1128 GB | 374/374 | 35,328 GB/s |
Model counts use default quantization at coding workload settings. Multi-GPU scaling factor: 0.92× per additional GPU.
Upgrade paths
See what you unlock with more powerful hardware
升级选项
8× NVIDIA H200 141GBMulti-GPUUnlocks 21 additional models that do not fit on the current setup.
Lifts average decode speed across fitting models by about 138%.
NVLink gives this scale-out path a cleaner inter-GPU story than PCIe-only builds.
~$30,000 MSRP
NVIDIA B200 180GB下一步升级Unlocks 6 additional models that do not fit on the current setup.
Lifts average decode speed across fitting models by about 22%.
~$30,000 MSRP
NVIDIA GB200 192GBNVIDIA 升级Unlocks 6 additional models that do not fit on the current setup.
Lifts average decode speed across fitting models by about 22%.
~$60,000 MSRP
Unlocks 10 additional models that do not fit on the current setup.
~$20,000 MSRP
Unlocks 11 additional models that do not fit on the current setup.
Lifts average decode speed across fitting models by about 14%.
~$8,000 MSRP
NVIDIA H200 141GB (141 GB VRAM) can run these top models: Qwen 3.5 122B A10B (score: 98/100), Devstral 2 123B Instruct (score: 98/100), Mistral Small 4 119B (score: 97/100). See the full compatibility list above.
NVIDIA H200 141GB has 141 GB of VRAM available for AI model inference. This determines which models and quantization levels you can run locally.
Yes, NVIDIA H200 141GB is excellent for running LLMs locally with top compatibility scores above 80/100.
For coding on NVIDIA H200 141GB, we recommend Qwen 3.5 122B A10B. It achieves 162.1 tokens per second with 131K context window. Qwen 3.5 122B A10B is a specialized fit for Coding. It is a recent-generation family, which helps on current local SOTA workloads. It fits natively with comfortable headroom. Context coverage stays within the requested workload envelope. Known distribution channels: huggingface, lm-studio.
There are 5 upgrade path(s) from NVIDIA H200 141GB: NVIDIA H200 141GB, NVIDIA B200 180GB. Upgrading would unlock larger models and faster inference speeds.
Yes, NVIDIA H200 141GB with 141 GB of usable memory can run Flux.1 Dev at FP16 natively. Flux is a 12B parameter diffusion transformer that produces high-quality images. You can also run the Schnell variant for faster generation.
NVIDIA H200 141GB (141 GB VRAM) can handle various AI generation tasks beyond LLMs. For image generation, SDXL and Stable Diffusion 3.5 run well. Flux.1 Dev also runs natively for state-of-the-art image quality. For video, LTX Video 2.3 can generate short clips. Check the AI Capability Matrix above for detailed compatibility.
NVIDIA H200 141GB is excellent for AI image generation. With 141 GB of usable memory, it runs all major diffusion models including Flux.1, SDXL, and Stable Diffusion 3.5 at full precision. You can generate high-resolution images quickly and even handle video generation models.
Yes, NVIDIA H200 141GB with 141 GB of usable memory can run Qwen 3.5 27B at Q8 (near-lossless, ~28.9 GB) or even FP16 (~55.4 GB) depending on your context needs. This setup provides an excellent experience with this model. Use Ollama or vLLM for best results.
With 141 GB VRAM on NVIDIA H200 141GB, use Q8_0 for most models — it is near-lossless and you have the memory for it. For 70B+ models, Q6_K offers excellent quality. Reserve Q4_K_M for 100B+ models or when you need maximum context length.
NVIDIA H200 141GB already has strong memory bandwidth, so the next limit is often memory capacity and context headroom rather than raw decode speed. For local LLMs, fit first and bandwidth second is the right mental model.
NVIDIA H200 141GB supports up to 8× GPU scaling via NVLink at 900 GB/s. With 8× GPUs, you get 1128 GB effective memory with a 0.92× scaling factor per GPU. This enables running models like Qwen 3.5 397B A17B and Kimi K2.5 that don't fit on a single card.
NVLink is recommended for NVIDIA H200 141GB multi-GPU inference, providing 900 GB/s interconnect bandwidth with only 8% scaling overhead. PCIe-only setups work but have higher overhead (~25%) due to limited inter-GPU bandwidth.
Compare with similar
