NVIDIA H20 (96GB)
The NVIDIA H20 (96GB) features 96 GB of dedicated VRAM and 4000 GB/s memory bandwidth, supporting CUDA and TensorRT-LLM runtimes for full-precision and quantized inference.
Memory
96 GB
Usable Memory Ceiling
96 GB
Bandwidth
4000 GB/s
Max Inference Class
70B+ Models (Q4)
TDP
400W
VRAM Scaling by Context Length
Estimated memory requirements as context length increases. Curves crossing above the reference line indicate Out of Memory (OOM).
Precision:
Best Models to Run on NVIDIA H20 (96GB)
Inference throughput, memory compatibility, and generation speeds for open-weights LLMs on NVIDIA H20 (96GB).
Sort by:
Context:
KV Cache:
GPUs:
| Rank | Model | Parameters | Precision | TPS (Approx.) | TTFT | Can Run |
|---|---|---|---|---|---|---|
#36 | Qwen 3.8 27B | 27B | INT4 Q8 FP16 | 158 tok/s 99 tok/s 54 tok/s | ~2.3s ~2.3s ~2.3s | |
#42 | Qwen 3.8 Flash | 176B (6B active) | INT4 | 373 tok/s | ~528 ms | |
#69 | Sarvam-105B | 106B (10.3B active) | INT4 | 294 tok/s | ~1.2s | |
#80 | K2 Horizon MoVA 36B A4B | 36B (4B active) | INT4 Q8 FP16 | 410 tok/s 334 tok/s 236 tok/s | ~442 ms ~443 ms ~444 ms | |
#83 | Ling 3.0 Flash VL | 124B (5.5B active) | INT4 | 386 tok/s | ~822 ms | |
#86 | Qwen3.5-27B | 27B | INT4 Q8 FP16 | 158 tok/s 99 tok/s 54 tok/s | ~2.3s ~2.3s ~2.3s | |
#88 | Qwen3.5-122B-A10B | 122B (10B active) | INT4 | 249 tok/s | ~1.2s | |
#93 | Sarvam-30B | 32B (2.4B active) | INT4 Q8 FP16 | 464 tok/s 404 tok/s 310 tok/s | ~309 ms ~309 ms ~310 ms | |
#99 | Muse Glimmer 30B | 30B | INT4 Q8 FP16 | 146 tok/s 90 tok/s 49 tok/s | ~2.5s ~2.5s ~2.5s | |
#102 | Phi-4 Reasoning Plus | 14B | INT4 Q8 FP16 | 197 tok/s 143 tok/s 88 tok/s | ~1.2s ~1.2s ~1.2s | |
#103 | Qwen3.6 35B A3B | 35B (3B active) | INT4 Q8 FP16 | 442 tok/s 374 tok/s 277 tok/s | ~360 ms ~361 ms ~362 ms | |
#105 | MiMo V2 Flash | 15B (309B active) | INT4 Q8 FP16 | 18 tok/s 10 tok/s 5 tok/s | ~24.6s ~24.6s ~24.7s | |
#110 | Gemma 4 26B A4B | 25.2B (3.8B active) | INT4 Q8 FP16 | 424 tok/s 347 tok/s 246 tok/s | ~405 ms ~405 ms ~406 ms | |
#113 | GLM-4.5-Air | 106B (12B active) | INT4 | 108 tok/s | ~1.3s | |
#117 | Nemotron 3.5 Bolt 74B A9B | 74B | INT4 Q8 | 69 tok/s 38 tok/s | ~6.1s ~6.6s | |
#120 | Qwen3.5-35B-A3B | 35B (3B active) | INT4 Q8 FP16 | 442 tok/s 374 tok/s 277 tok/s | ~360 ms ~361 ms ~362 ms | |
#121 | Qwen3.5-9B | 9B | INT4 Q8 FP16 | 319 tok/s 227 tok/s 138 tok/s | ~768 ms ~769 ms ~772 ms | |
#122 | Ling 3.0 Flash | 124B (5.1B active) | INT4 | 397 tok/s | ~790 ms | |
#127 | K2 Horizon 7B | 7B | INT4 Q8 FP16 | 360 tok/s 265 tok/s 168 tok/s | ~601 ms ~602 ms ~605 ms | |
#130 | Gemma 4 12B | 11.95B | INT4 Q8 FP16 | 217 tok/s 161 tok/s 100 tok/s | ~1.0s ~1.0s ~1.0s | |
#131 | Qwen3 Next 80B A3B | 80B (79B active) | INT4 Q8 | 52 tok/s 30 tok/s | ~6.5s ~7.7s | |
#135 | GPT-OSS 120B | 117B (5.1B active) | INT4 | 109 tok/s | ~853 ms | |
#136 | Qwen3.5-4B | 4B | INT4 Q8 FP16 | 445 tok/s 355 tok/s 246 tok/s | ~353 ms ~354 ms ~355 ms | |
#139 | GPT-OSS 20B | 21B (3.6B active) | INT4 Q8 FP16 | 275 tok/s 248 tok/s 194 tok/s | ~380 ms ~380 ms ~381 ms | |
#140 | GLM-4V | 9B | INT4 Q8 FP16 | 257 tok/s 196 tok/s 126 tok/s | ~769 ms ~770 ms ~773 ms | |
#145 | NVIDIA Nemotron 3 Nano 30B-A3B | 3.5B (30B active) | INT4 Q8 FP16 | 140 tok/s 88 tok/s 48 tok/s | ~2.4s ~2.4s ~2.4s | |
#147 | Nemotron 3.5 Super VL | 120B | INT4 | 44 tok/s | ~9.9s | |
#147 | Mistral Medium 3.5 128B | 128B | INT4 | 32 tok/s | ~11.5s | |
#150 | Qwen3-30B-A3B | 30B (3B active) | INT4 Q8 FP16 | 445 tok/s 376 tok/s 278 tok/s | ~353 ms ~354 ms ~355 ms | |
#152 | Granite 4.2 8B | 8B | INT4 Q8 FP16 | 338 tok/s 244 tok/s 152 tok/s | ~684 ms ~685 ms ~688 ms | |
#157 | Gemma 4 E4B | 8B | INT4 Q8 FP16 | 338 tok/s 244 tok/s 152 tok/s | ~684 ms ~685 ms ~688 ms | |
#158 | Mistral Large 3 | 41B (675B active) | INT4 Q8 FP16 | 8 tok/s 5 tok/s 2 tok/s | ~53.7s ~53.8s ~63.5s | |
#159 | Qwen2.5-72B | 72B | INT4 Q8 | 70 tok/s 41 tok/s | ~5.9s ~5.9s | |
#161 | GLM-4 | 32B | INT4 Q8 FP16 | 112 tok/s 75 tok/s 43 tok/s | ~2.7s ~2.7s ~2.7s | |
#162 | DeepSeek-R1 70B | 70B | INT4 Q8 | 74 tok/s 43 tok/s | ~5.8s ~5.8s | |
#163 | Gemma 3 27B | 27B | INT4 Q8 FP16 | 158 tok/s 99 tok/s 54 tok/s | ~2.3s ~2.3s ~2.3s | |
#164 | Phi-4 | 14B | INT4 Q8 FP16 | 248 tok/s 166 tok/s 96 tok/s | ~1.2s ~1.2s ~1.2s | |
#165 | Llama 3.3 70B | 70B | INT4 Q8 | 72 tok/s 42 tok/s | ~5.8s ~5.8s | |
#166 | Qwen3-32B | 32B | INT4 Q8 FP16 | 139 tok/s 85 tok/s 46 tok/s | ~2.7s ~2.7s ~2.7s | |
#167 | Hunyuan Standard | 52B (389B active) | INT4 Q8 | 14 tok/s 8 tok/s | ~31.0s ~31.1s | |
#168 | DeepSeek-R1 32B | 32B | INT4 Q8 FP16 | 143 tok/s 86 tok/s 46 tok/s | ~2.7s ~2.7s ~2.7s | |
#169 | Gemma 4 31B | 30.7B | INT4 Q8 FP16 | 144 tok/s 88 tok/s 48 tok/s | ~2.6s ~2.6s ~2.6s | |
#171 | Magistral Small | 24B | INT4 Q8 FP16 | 138 tok/s 95 tok/s 55 tok/s | ~2.0s ~2.0s ~2.0s | |
#172 | Qwen2.5-32B | 32B | INT4 Q8 FP16 | 139 tok/s 85 tok/s 46 tok/s | ~2.7s ~2.7s ~2.7s | |
#174 | OLMo 3.1 32B Instruct | 32B | INT4 Q8 FP16 | 112 tok/s 75 tok/s 43 tok/s | ~2.7s ~2.7s ~2.7s | |
#175 | Llama 3.1 70B | 70B | INT4 Q8 | 72 tok/s 42 tok/s | ~5.8s ~5.8s | |
#176 | DeepSeek-R1 14B | 14B | INT4 Q8 FP16 | 254 tok/s 168 tok/s 97 tok/s | ~1.2s ~1.2s ~1.2s | |
#177 | Ministral 3 14B | 14B | INT4 Q8 FP16 | 197 tok/s 143 tok/s 88 tok/s | ~1.2s ~1.2s ~1.2s | |
#178 | Mistral-Large-2407 | 123B | INT4 | 43 tok/s | ~10.1s | |
#179 | Llama 4 Scout | 109B (17B active) | INT4 | 195 tok/s | ~1.7s | |
#180 | Gemma 3 12B | 12B | INT4 Q8 FP16 | 272 tok/s 186 tok/s 110 tok/s | ~1.0s ~1.0s ~1.0s | |
#182 | Command A | 111B | INT4 | 38 tok/s | ~9.2s | |
#183 | Qwen2-72B | 72B | INT4 Q8 | 70 tok/s 41 tok/s | ~5.9s ~5.9s | |
#184 | Mistral-Small-2501 | 24B | INT4 Q8 FP16 | 173 tok/s 109 tok/s 60 tok/s | ~2.0s ~2.0s ~2.0s | |
#185 | Ministral 3 8B | 8B | INT4 Q8 FP16 | 275 tok/s 212 tok/s 138 tok/s | ~685 ms ~686 ms ~688 ms | |
#186 | Gemma 2 27B | 27B | INT4 Q8 FP16 | 158 tok/s 99 tok/s 54 tok/s | ~2.3s ~2.3s ~2.3s | |
#187 | OLMo 3 7B Think | 7B | INT4 Q8 FP16 | 293 tok/s 229 tok/s 153 tok/s | ~602 ms ~603 ms ~605 ms | |
#188 | Qwen3.5-2B | 2B | INT4 Q8 FP16 | 527 tok/s 459 tok/s 357 tok/s | ~189 ms ~189 ms ~190 ms | |
#189 | MiniCPM5-2B | 2B | INT4 Q8 FP16 | 527 tok/s 459 tok/s 357 tok/s | ~189 ms ~189 ms ~190 ms | |
#190 | Qwen2.5-14B | 14B | INT4 Q8 FP16 | 248 tok/s 166 tok/s 96 tok/s | ~1.2s ~1.2s ~1.2s | |
#191 | OLMo 3.1 32B Think | 32B | INT4 Q8 FP16 | 112 tok/s 75 tok/s 43 tok/s | ~2.7s ~2.7s ~2.7s | |
#192 | Llama 3 70B | 70B | INT4 Q8 | 72 tok/s 42 tok/s | ~5.8s ~5.8s | |
#193 | Gemma 4 E2B | 5.1B | INT4 Q8 FP16 | 409 tok/s 316 tok/s 210 tok/s | ~444 ms ~445 ms ~447 ms | |
#194 | Command R Plus | 104B | INT4 | 40 tok/s | ~8.6s | |
#195 | Ministral 3 3B | 3B | INT4 Q8 FP16 | 398 tok/s 345 tok/s 261 tok/s | ~271 ms ~272 ms ~273 ms | |
#196 | DeepSeek-R1 8B | 8B | INT4 Q8 FP16 | 343 tok/s 247 tok/s 153 tok/s | ~684 ms ~685 ms ~687 ms | |
#198 | Gemma 2 9B | 9B | INT4 Q8 FP16 | 319 tok/s 227 tok/s 138 tok/s | ~768 ms ~769 ms ~772 ms | |
#199 | Nemotron 3.5 Lightning 30B A3B | 30B (3B active) | INT4 Q8 FP16 | 445 tok/s 376 tok/s 278 tok/s | ~353 ms ~354 ms ~355 ms | |
#200 | Kimi Linear 48B A3B Instruct | 48B (3B active) | INT4 Q8 | 227 tok/s 215 tok/s | ~383 ms ~384 ms | |
#201 | Qwen3-4B | 4B | INT4 Q8 FP16 | 445 tok/s 355 tok/s 246 tok/s | ~353 ms ~354 ms ~355 ms | |
#202 | Qwen3-14B | 14B | INT4 Q8 FP16 | 248 tok/s 166 tok/s 96 tok/s | ~1.2s ~1.2s ~1.2s | |
#203 | Gemma 3 4B | 4B | INT4 Q8 FP16 | 445 tok/s 355 tok/s 246 tok/s | ~353 ms ~354 ms ~355 ms | |
#204 | DeepSeek-R1 7B | 7B | INT4 Q8 FP16 | 365 tok/s 268 tok/s 169 tok/s | ~601 ms ~602 ms ~605 ms | |
#205 | OLMo 3 32B Base | 32B | INT4 Q8 FP16 | 112 tok/s 75 tok/s 43 tok/s | ~2.7s ~2.7s ~2.7s | |
#206 | Qwen3.5-0.8B | 0.8B | INT4 Q8 FP16 | 602 tok/s 564 tok/s 494 tok/s | ~78 ms ~78 ms ~79 ms | |
#207 | Llama 3.1 8B | 8B | INT4 Q8 FP16 | 338 tok/s 244 tok/s 152 tok/s | ~684 ms ~685 ms ~688 ms | |
#207 | Qwen3-8B | 8B | INT4 Q8 FP16 | 338 tok/s 244 tok/s 152 tok/s | ~684 ms ~685 ms ~688 ms | |
#209 | Mixtral-8x22B-v0.1 | 176B (22B active) | INT4 | 149 tok/s | ~2.5s | |
#210 | Nemotron 3.5 Lightning | 30B (3B active) | INT4 Q8 FP16 | 255 tok/s 238 tok/s 195 tok/s | ~355 ms ~355 ms ~356 ms | |
#211 | Command R | 35B | INT4 Q8 FP16 | 105 tok/s 70 tok/s 38 tok/s | ~2.9s ~2.9s ~3.2s | |
#212 | Llama 3 8B | 8B | INT4 Q8 FP16 | 338 tok/s 244 tok/s 152 tok/s | ~684 ms ~685 ms ~688 ms | |
#213 | Qwen3-1.7B | 1.7B | INT4 Q8 FP16 | 544 tok/s 481 tok/s 383 tok/s | ~162 ms ~162 ms ~163 ms | |
#214 | Gemma 3 270M | 0.27B | INT4 Q8 FP16 | 628 tok/s 615 tok/s 584 tok/s | ~28 ms ~28 ms ~28 ms | |
#216 | Qwen3-0.6B | 0.6B | INT4 Q8 FP16 | 618 tok/s 587 tok/s 529 tok/s | ~59 ms ~59 ms ~59 ms | |
#217 | Phi-4-Mini | 3.8B | INT4 Q8 FP16 | 452 tok/s 364 tok/s 254 tok/s | ~337 ms ~337 ms ~339 ms | |
#218 | Qwen2.5-7B | 7B | INT4 Q8 FP16 | 360 tok/s 265 tok/s 168 tok/s | ~601 ms ~602 ms ~605 ms | |
#219 | Mixtral-8x7B-v0.1 | 46.7B (7B active) | INT4 Q8 FP16 | 335 tok/s 252 tok/s 146 tok/s | ~698 ms ~699 ms ~824 ms | |
#220 | Ministral-8B-2410 | 8B | INT4 Q8 FP16 | 338 tok/s 244 tok/s 152 tok/s | ~684 ms ~685 ms ~688 ms | |
#221 | DeepSeek-R1 1.5B | 1.5B | INT4 Q8 FP16 | 556 tok/s 498 tok/s 403 tok/s | ~143 ms ~144 ms ~144 ms | |
#222 | Llama 3.2 3B | 3B | INT4 Q8 FP16 | 483 tok/s 401 tok/s 291 tok/s | ~271 ms ~271 ms ~272 ms | |
#223 | Phi-3-mini | 3.8B | INT4 Q8 FP16 | 452 tok/s 364 tok/s 254 tok/s | ~337 ms ~337 ms ~339 ms | |
#224 | Gemma 3n E2B IT | 6B (2B active) | INT4 Q8 FP16 | 398 tok/s 364 tok/s 297 tok/s | ~202 ms ~202 ms ~203 ms | |
#225 | Gemma 2 2B | 2B | INT4 Q8 FP16 | 527 tok/s 459 tok/s 357 tok/s | ~189 ms ~189 ms ~190 ms | |
#226 | Mistral-7B-Instruct-v0.1 | 7.3B | INT4 Q8 FP16 | 353 tok/s 259 tok/s 163 tok/s | ~626 ms ~627 ms ~629 ms | |
#227 | Llama 3.2 1B | 1B | INT4 Q8 FP16 | 587 tok/s 542 tok/s 464 tok/s | ~99 ms ~99 ms ~99 ms | |
#228 | Phi-3-medium | 14B | INT4 Q8 FP16 | 248 tok/s 166 tok/s 96 tok/s | ~1.2s ~1.2s ~1.2s | |
#229 | Yi-34B | 34B | INT4 Q8 FP16 | 108 tok/s 71 tok/s 41 tok/s | ~2.8s ~2.8s ~2.8s | |
#230 | OLMo 3 7B Instruct | 7B | INT4 Q8 FP16 | 293 tok/s 229 tok/s 153 tok/s | ~602 ms ~603 ms ~605 ms | |
#231 | Qwen2-7B | 7B | INT4 Q8 FP16 | 360 tok/s 265 tok/s 168 tok/s | ~601 ms ~602 ms ~605 ms | |
#232 | Phi-3-small | 7B | INT4 Q8 FP16 | 360 tok/s 265 tok/s 168 tok/s | ~601 ms ~602 ms ~605 ms | |
#233 | Mistral-7B-Instruct-v0.2 | 7.3B | INT4 Q8 FP16 | 353 tok/s 259 tok/s 163 tok/s | ~626 ms ~627 ms ~629 ms | |
#234 | Gemma 3 1B | 1B | INT4 Q8 FP16 | 587 tok/s 542 tok/s 464 tok/s | ~99 ms ~99 ms ~99 ms | |
#235 | Falcon-180B | 180B | INT4 | 28 tok/s | ~17.3s | |
#236 | Gemma 1 7B | 7B | INT4 Q8 FP16 | 293 tok/s 229 tok/s 153 tok/s | ~602 ms ~603 ms ~605 ms | |
#237 | Gemma 1 2B | 2B | INT4 Q8 FP16 | 540 tok/s 468 tok/s 362 tok/s | ~189 ms ~189 ms ~190 ms | |
#238 | ChatGLM3-6B | 6B | INT4 Q8 FP16 | 313 tok/s 250 tok/s 170 tok/s | ~519 ms ~520 ms ~522 ms | |
#239 | ChatGLM2-6B | 6B | INT4 Q8 FP16 | 313 tok/s 250 tok/s 170 tok/s | ~519 ms ~520 ms ~522 ms | |
#240 | ChatGLM-6B | 6B | INT4 Q8 FP16 | 313 tok/s 250 tok/s 170 tok/s | ~519 ms ~520 ms ~522 ms | |
Unranked Models (Sorted by Speed) | ||||||
- | DINOv3 ViT-L/16 | 0.3B | INT4 Q8 FP16 | 642 tok/s 625 tok/s 590 tok/s | ~31 ms ~31 ms ~31 ms | |
- | Qwen2-0.5B | 0.5B | INT4 Q8 FP16 | 626 tok/s 600 tok/s 548 tok/s | ~49 ms ~49 ms ~49 ms | |
- | Qwen2.5-0.5B | 0.5B | INT4 Q8 FP16 | 626 tok/s 600 tok/s 548 tok/s | ~49 ms ~49 ms ~49 ms | |
- | ERNIE-4.5-0.3B | 0.3B | INT4 Q8 FP16 | 623 tok/s 608 tok/s 576 tok/s | ~31 ms ~31 ms ~31 ms | |
- | ERNIE-4.5-0.3B-Base | 0.3B | INT4 Q8 FP16 | 623 tok/s 608 tok/s 576 tok/s | ~31 ms ~31 ms ~31 ms | |
- | Liquid D1-Omni 600M | 0.6B | INT4 Q8 FP16 | 618 tok/s 587 tok/s 529 tok/s | ~59 ms ~59 ms ~59 ms | |
- | Falcon-1B | 1B | INT4 Q8 FP16 | 596 tok/s 549 tok/s 469 tok/s | ~99 ms ~99 ms ~99 ms | |
- | Falcon3-1B | 1B | INT4 Q8 FP16 | 587 tok/s 542 tok/s 464 tok/s | ~99 ms ~99 ms ~99 ms | |
- | MiniCPM-V 4.7 1B | 1B | INT4 Q8 FP16 | 587 tok/s 542 tok/s 464 tok/s | ~99 ms ~99 ms ~99 ms | |
- | Qwen2-1.5B | 1.5B | INT4 Q8 FP16 | 556 tok/s 498 tok/s 403 tok/s | ~143 ms ~144 ms ~144 ms | |
- | Qwen2.5-1.5B | 1.5B | INT4 Q8 FP16 | 556 tok/s 498 tok/s 403 tok/s | ~143 ms ~144 ms ~144 ms | |
- | Spark X2.5 1.7B | 1.7B | INT4 Q8 FP16 | 544 tok/s 481 tok/s 383 tok/s | ~162 ms ~162 ms ~163 ms | |
- | Hy-MT2 1.8B | 1.8B | INT4 Q8 FP16 | 539 tok/s 474 tok/s 374 tok/s | ~170 ms ~170 ms ~171 ms | |
- | Falcon-3B | 3B | INT4 Q8 FP16 | 494 tok/s 408 tok/s 295 tok/s | ~271 ms ~271 ms ~272 ms | |
- | CroissantLLM Base | 1.3B | INT4 Q8 FP16 | 490 tok/s 455 tok/s 384 tok/s | ~125 ms ~126 ms ~126 ms | |
- | Phi-1 | 1.3B | INT4 Q8 FP16 | 490 tok/s 455 tok/s 384 tok/s | ~125 ms ~126 ms ~126 ms | |
- | Phi-1.5 | 1.3B | INT4 Q8 FP16 | 490 tok/s 455 tok/s 384 tok/s | ~125 ms ~126 ms ~126 ms | |
- | Mellum 2.1 12B A2.5B Thinking | 12B (2.5B active) | INT4 Q8 FP16 | 486 tok/s 416 tok/s 314 tok/s | ~265 ms ~265 ms ~266 ms | |
- | DeepSeek-R1 3B | 3B | INT4 Q8 FP16 | 485 tok/s 402 tok/s 292 tok/s | ~271 ms ~271 ms ~272 ms | |
- | Falcon3-3B | 3B | INT4 Q8 FP16 | 483 tok/s 401 tok/s 291 tok/s | ~271 ms ~271 ms ~272 ms | |
- | LFM2-VL 3B | 3B | INT4 Q8 FP16 | 483 tok/s 401 tok/s 291 tok/s | ~271 ms ~271 ms ~272 ms | |
- | Liquid D1 3B | 3B | INT4 Q8 FP16 | 483 tok/s 401 tok/s 291 tok/s | ~271 ms ~271 ms ~272 ms | |
- | Liquid D1-VL 3B | 3B | INT4 Q8 FP16 | 483 tok/s 401 tok/s 291 tok/s | ~271 ms ~271 ms ~272 ms | |
- | Liquid LFM2-VL 3B D1 | 3B | INT4 Q8 FP16 | 483 tok/s 401 tok/s 291 tok/s | ~271 ms ~271 ms ~272 ms | |
- | Qwen2.5-3B | 3B | INT4 Q8 FP16 | 483 tok/s 401 tok/s 291 tok/s | ~271 ms ~271 ms ~272 ms | |
- | LFM2.5 24B A2B | 24B (2.5B active) | INT4 Q8 FP16 | 469 tok/s 404 tok/s 307 tok/s | ~300 ms ~300 ms ~301 ms | |
- | ERNIE-4.5-21B-A3B | 21B (3B active) | INT4 Q8 FP16 | 454 tok/s 383 tok/s 282 tok/s | ~331 ms ~331 ms ~332 ms | |
- | ERNIE-4.5-21B-A3B-Base | 21B (3B active) | INT4 Q8 FP16 | 454 tok/s 383 tok/s 282 tok/s | ~331 ms ~331 ms ~332 ms | |
- | ERNIE-4.5-VL-28B-A3B | 28B (3B active) | INT4 Q8 FP16 | 447 tok/s 378 tok/s 279 tok/s | ~348 ms ~348 ms ~349 ms | |
- | ERNIE-4.5-VL-28B-A3B-Base | 28B (3B active) | INT4 Q8 FP16 | 447 tok/s 378 tok/s 279 tok/s | ~348 ms ~348 ms ~349 ms | |
- | FrogNano 4B | 4B | INT4 Q8 FP16 | 445 tok/s 355 tok/s 246 tok/s | ~353 ms ~354 ms ~355 ms | |
- | Hy-Memory 30B A3B | 30B (3B active) | INT4 Q8 FP16 | 445 tok/s 376 tok/s 278 tok/s | ~353 ms ~354 ms ~355 ms | |
- | Hy-MT2 30B A3B | 30B (3B active) | INT4 Q8 FP16 | 445 tok/s 376 tok/s 278 tok/s | ~353 ms ~354 ms ~355 ms | |
- | Spark X2.5 4B | 4B | INT4 Q8 FP16 | 445 tok/s 355 tok/s 246 tok/s | ~353 ms ~354 ms ~355 ms | |
- | MiniCPM-V 4.7 35B A3B | 35B (3B active) | INT4 Q8 FP16 | 442 tok/s 374 tok/s 277 tok/s | ~360 ms ~361 ms ~362 ms | |
- | EmbeddingGemma 2 | 2B | INT4 Q8 FP16 | 436 tok/s 393 tok/s 316 tok/s | ~189 ms ~189 ms ~190 ms | |
- | Phi-2 | 2.7B | INT4 Q8 FP16 | 409 tok/s 358 tok/s 275 tok/s | ~247 ms ~247 ms ~248 ms | |
- | MaLLaM-3B | 3B | INT4 Q8 FP16 | 398 tok/s 345 tok/s 261 tok/s | ~271 ms ~272 ms ~273 ms | |
- | SmolLM3 3B | 3B | INT4 Q8 FP16 | 398 tok/s 345 tok/s 261 tok/s | ~271 ms ~272 ms ~273 ms | |
- | AliceAI Foundation 80B A3B Base | 80B (3B active) | INT4 Q8 | 395 tok/s 326 tok/s | ~484 ms ~527 ms | |
- | Kolibri 1 | 78B (3.5B active) | INT4 Q8 | 384 tok/s 310 tok/s | ~520 ms ~565 ms | |
- | Falcon-7B | 7B | INT4 Q8 FP16 | 369 tok/s 270 tok/s 170 tok/s | ~601 ms ~602 ms ~605 ms | |
- | Falcon3-7B | 7B | INT4 Q8 FP16 | 360 tok/s 265 tok/s 168 tok/s | ~601 ms ~602 ms ~605 ms | |
- | Hy-MT2 7B | 7B | INT4 Q8 FP16 | 360 tok/s 265 tok/s 168 tok/s | ~601 ms ~602 ms ~605 ms | |
- | Mistral-7B-v0.1 | 7.3B | INT4 Q8 FP16 | 353 tok/s 259 tok/s 163 tok/s | ~626 ms ~627 ms ~629 ms | |
- | Youtu Parsing Omni | 8B | INT4 Q8 FP16 | 343 tok/s 247 tok/s 153 tok/s | ~684 ms ~685 ms ~687 ms | |
- | LensVLM 9B | 9B | INT4 Q8 FP16 | 319 tok/s 227 tok/s 138 tok/s | ~768 ms ~769 ms ~772 ms | |
- | MiMo V2.6 Distill Qwen 9B | 9B | INT4 Q8 FP16 | 319 tok/s 227 tok/s 138 tok/s | ~768 ms ~769 ms ~772 ms | |
- | ChatGLM3-6B-32K | 6B | INT4 Q8 FP16 | 313 tok/s 250 tok/s 170 tok/s | ~519 ms ~520 ms ~522 ms | |
- | Yi-6B | 6B | INT4 Q8 FP16 | 313 tok/s 250 tok/s 170 tok/s | ~519 ms ~520 ms ~522 ms | |
- | Falcon3-10B | 10B | INT4 Q8 FP16 | 302 tok/s 211 tok/s 127 tok/s | ~851 ms ~852 ms ~855 ms | |
- | Kimi-VL-A3B-Instruct | 16B (3B active) | INT4 Q8 FP16 | 302 tok/s 275 tok/s 219 tok/s | ~318 ms ~319 ms ~320 ms | |
- | Falcon2-11B | 11B | INT4 Q8 FP16 | 294 tok/s 201 tok/s 119 tok/s | ~934 ms ~935 ms ~939 ms | |
- | Hunyuan Lite | 7B | INT4 Q8 FP16 | 293 tok/s 229 tok/s 153 tok/s | ~602 ms ~603 ms ~605 ms | |
- | OLMo 3 7B Base | 7B | INT4 Q8 FP16 | 293 tok/s 229 tok/s 153 tok/s | ~602 ms ~603 ms ~605 ms | |
- | SEA-LION-7B | 7.1B | INT4 Q8 FP16 | 291 tok/s 228 tok/s 151 tok/s | ~610 ms ~611 ms ~613 ms | |
- | SEA-LION-7B-Instruct | 7.1B | INT4 Q8 FP16 | 291 tok/s 228 tok/s 151 tok/s | ~610 ms ~611 ms ~613 ms | |
- | Sahabat-AI-Llama3-8B-Instruct | 8B | INT4 Q8 FP16 | 275 tok/s 212 tok/s 138 tok/s | ~685 ms ~686 ms ~688 ms | |
- | Typhoon-2-8B | 8B | INT4 Q8 FP16 | 275 tok/s 212 tok/s 138 tok/s | ~685 ms ~686 ms ~688 ms | |
- | Laguna S 2.1 | 118B (8B active) | INT4 | 273 tok/s | ~1.0s | |
- | GLM-4-9B | 9B | INT4 Q8 FP16 | 257 tok/s 196 tok/s 126 tok/s | ~769 ms ~770 ms ~773 ms | |
- | GLM-4-9B-Chat | 9B | INT4 Q8 FP16 | 257 tok/s 196 tok/s 126 tok/s | ~769 ms ~770 ms ~773 ms | |
- | GLM-4-9B-Chat-1M | 9B | INT4 Q8 FP16 | 257 tok/s 196 tok/s 126 tok/s | ~769 ms ~770 ms ~773 ms | |
- | Yi-9B | 9B | INT4 Q8 FP16 | 257 tok/s 196 tok/s 126 tok/s | ~769 ms ~770 ms ~773 ms | |
- | Sahabat-AI-Gemma2-9B | 9.2B | INT4 Q8 FP16 | 254 tok/s 193 tok/s 124 tok/s | ~785 ms ~786 ms ~789 ms | |
- | Sahabat-AI-Gemma2-9B-Instruct | 9.2B | INT4 Q8 FP16 | 254 tok/s 193 tok/s 124 tok/s | ~785 ms ~786 ms ~789 ms | |
- | GreenMind-14B-R1 | 14B | INT4 Q8 FP16 | 197 tok/s 143 tok/s 88 tok/s | ~1.2s ~1.2s ~1.2s | |
- | AREX-2 | 27B | INT4 Q8 FP16 | 158 tok/s 99 tok/s 54 tok/s | ~2.3s ~2.3s ~2.3s | |
- | PPLX Decider v1 27B | 27B | INT4 Q8 FP16 | 158 tok/s 99 tok/s 54 tok/s | ~2.3s ~2.3s ~2.3s | |
- | PPLX Decider v1.1 27B | 27B | INT4 Q8 FP16 | 158 tok/s 99 tok/s 54 tok/s | ~2.3s ~2.3s ~2.3s | |
- | Agnes 3.0 Flash | 33B | INT4 Q8 FP16 | 136 tok/s 83 tok/s 45 tok/s | ~2.7s ~2.7s ~2.8s | |
- | Hunyuan A13B | 80B (13B active) | INT4 Q8 | 127 tok/s 96 tok/s | ~1.3s ~1.5s | |
- | Falcon-40B | 40B | INT4 Q8 FP16 | 120 tok/s 71 tok/s 36 tok/s | ~3.3s ~3.3s ~3.6s | |
- | Typhoon-2-70B | 70B | INT4 Q8 | 59 tok/s 36 tok/s | ~5.8s ~6.3s | |
- | Kimi-Dev-72B | 72B | INT4 Q8 | 57 tok/s 35 tok/s | ~5.9s ~6.5s | |
- | GLM-130B | 130B | INT4 | 32 tok/s | ~11.6s | |
Inference Capacity
Maximum model parameter class runnable by weight precision.
Context:
4-bit
70B+ Models
8-bit
70B Models
16-bit
32B Models
Fine-Tuning Capacity
Hardware capacity for local training and adapter fine-tuning.
Context Length:
QLoRA
70B+ Models
LoRA
70B Models
Full Parameter Training
7B-8B Models
Workload Recommendations
Guidance for optimal precision and operational ceilings on NVIDIA H20 (96GB).
Inference Sweet Spot
Suitable for 70B parameter models at Q4/Q8 with extended 32k context windows, or high-throughput batching of smaller dense architectures.
Bandwidth & Speed Profile
With 4000 GB/s aggregate bandwidth, batch size 1 inference operates in a memory-bandwidth bound regime. Generates approximately 889 tok/s on an 8B Q4 model and 105 tok/s on a 70B Q4 model.
Frequently Asked Questions
©2025 ApX Machine Learning