AI Index

Open-weight models that run on NVIDIA DGX Spark · 128 GB

  • Unified memory 128 GB
  • Bandwidth 273 GB/s
  • Usable 96.0 GB

Specs: Memory · Personal AI Supercomputer Powered by Blackwell | NVIDIA DGX Spark · Bandwidth · Personal AI Supercomputer Powered by Blackwell | NVIDIA DGX Spark

FitModelParameters / activePrecisionMemoryEst. speedLongest contextRuns as
Runs wellMing-3.0-audio-flash Ant Group128.2B / 6.5BQ4_K_M76.5 / 96.0 GB31–47 tok/s128KUnified memory
Runs wellLing-3.0-flash Ant Group127.5B / 5.1BQ5_086.1 / 96.0 GB35–52 tok/s223KUnified memory
Runs wellQwen3.5-122B-A10B Alibaba Qwen125.1B / 10BQ5_084.3 / 96.0 GB19–29 tok/s256KUnified memory
Runs wellLing-3.0-flash-VL Ant Group124.8B / 5.5BQ5_084.3 / 96.0 GB33–49 tok/s128KUnified memory
Runs wellgpt-oss-120b OpenAI116.8B / 5.1BMXFP4 (native)maker’s file64.1 / 96.0 GB43–65 tok/s128KUnified memory
Runs wellGLM-4.5-Air Zhipu AI110.5BIQ4_XS58.8 / 96.0 GB13–19 tok/s125KUnified memory
Runs wellGLM-4.5V Zhipu AI107.7B / 12BQ5_K_M76.4 / 96.0 GB14–20 tok/s63KUnified memory
Runs wellGLM-4.6V Zhipu AI107.7BQ4_K_M65.4 / 96.0 GB13–20 tok/s125KUnified memory
Runs wellLing-2.6-flash Ant Group107.5B / 7.4BQ5_K_M75.0 / 96.0 GB25–37 tok/s128KUnified memory
Runs wellQwen3-Next-80B-A3B-Instruct Alibaba Qwen81.3B / 3BQ8_0maker’s file83.1 / 96.0 GB41–61 tok/s256KUnified memory
Runs wellQwen3-Next-80B-A3B-Thinking Alibaba Qwen81.3B / 3BQ8_0maker’s file83.1 / 96.0 GB41–61 tok/s256KUnified memory
Runs wellQwen3-Coder-Next Alibaba Qwen79.7B / 3BQ8_0maker’s file83.1 / 96.0 GB40–60 tok/s256KUnified memory
Runs wellQwen3.5-35B-A3B Alibaba Qwen36B / 3BBF16 (native)maker’s file70.5 / 96.0 GB22–33 tok/s256KUnified memory
Runs wellQwen3.6-35B-A3B Alibaba Qwen36B / 3BBF16 (native)maker’s file70.5 / 96.0 GB22–33 tok/s256KUnified memory
Runs wellQwen3-Omni-30B-A3B-Captioner Alibaba Qwen31.7BBF16 (native)maker’s file62.8 / 96.0 GB14–21 tok/s64KUnified memory
Runs wellGLM-4.7-Flash Zhipu AI31.2BBF16 (native)maker’s file61.5 / 96.0 GB13–19 tok/s195KUnified memory
Runs wellQwen3-VL-30B-A3B-Instruct Alibaba Qwen31.1BBF16 (native)maker’s file61.5 / 96.0 GB16–24 tok/s128KUnified memory
Runs wellQwen3-VL-30B-A3B-Thinking Alibaba Qwen31.1BBF16 (native)maker’s file61.5 / 96.0 GB16–24 tok/s128KUnified memory
Runs wellQwen3-30B-A3B-Instruct-2507 Alibaba Qwen30.5B / 3.3BBF16 (native)maker’s file60.5 / 96.0 GB18–28 tok/s256KUnified memory
Runs wellQwen3-30B-A3B-Thinking-2507 Alibaba Qwen30.5B / 3.3BBF16 (native)maker’s file60.5 / 96.0 GB18–28 tok/s256KUnified memory
Runs wellQwen3-30B-A3B Alibaba Qwen30.5B / 3.3BBF16 (native)maker’s file60.5 / 96.0 GB18–28 tok/s128KUnified memory
Runs wellQwen3-Coder-30B-A3B-Instruct Alibaba Qwen30.5B / 3.3BBF16 (native)maker’s file60.5 / 96.0 GB18–28 tok/s256KUnified memory
Runs wellHy-Memory-30B-A3B Tencent30.1B / 3BBF16 (native)maker’s file59.6 / 96.0 GB20–30 tok/s63KUnified memory
Runs wellGemma 4 26B A4B IT Google25.8B / 3.8BBF16 (native)maker’s file51.0 / 96.0 GB17–25 tok/s250KUnified memory
Runs wellERNIE-4.5-21B-A3B-Thinking Baidu21.8B / 3BBF16 (native)maker’s file43.1 / 96.0 GB21–32 tok/s128KUnified memory
Runs wellGPT OSS 20B OpenAI20.9B / 3.6BMXFP4 (native)maker’s file13.6 / 96.0 GB53–80 tok/s128KUnified memory
Runs wellQwen2.5-14B-Instruct-1M Alibaba Qwen14.8BQ4_K_M10.4 / 96.0 GB13–19 tok/s464KUnified memory
Runs wellQwen2.5-14B-Instruct Alibaba Qwen14.8BQ4_K_Mmaker’s file10.4 / 96.0 GB13–19 tok/s32KUnified memory
Runs wellDeepSeek-R1-Distill-Qwen-14B DeepSeek14.8BQ4_K_M10.4 / 96.0 GB13–19 tok/s32KUnified memory
Runs wellQwen3-14B Alibaba Qwen14.8BQ4_K_Mmaker’s file10.1 / 96.0 GB13–20 tok/s128KUnified memory
Runs wellGemma 3 12B IT Google12.2BQ4_K_M10.2 / 96.0 GB13–20 tok/s128KUnified memory
Runs wellQwen2.5-Omni-7B Alibaba Qwen10.7BQ6_K9.1 / 96.0 GB15–22 tok/s32KUnified memory
Runs wellGLM-4.6V-Flash Zhipu AI10.3BFP810.4 / 96.0 GB13–19 tok/s125KUnified memory
Runs wellQwen3.5-9B Alibaba Qwen9.7BQ8_010.3 / 96.0 GB13–19 tok/s256KUnified memory
Runs wellQianfan-VL-8B Baidu8.8BQ8_010.2 / 96.0 GB13–20 tok/s32KUnified memory
Runs wellGranite Switch 4.1 8B IBM8.8BFP810.0 / 96.0 GB13–20 tok/s128KUnified memory
Runs wellQwen3-VL-8B-Instruct Alibaba Qwen8.8BQ8_0maker’s file9.7 / 96.0 GB14–21 tok/s128KUnified memory
Runs wellQwen3-VL-8B-Thinking Alibaba Qwen8.8BQ8_0maker’s file9.7 / 96.0 GB14–21 tok/s128KUnified memory
Runs wellpplx-embed-v2-context-9b Perplexity8.4BQ8_09.1 / 96.0 GB15–22 tok/s256KUnified memory
Runs wellQwen2.5-VL-7B-Instruct Alibaba Qwen8.3BQ8_09.1 / 96.0 GB15–22 tok/s125KUnified memory
Runs wellQwen3-8B Alibaba Qwen8.2BQ8_0maker’s file9.7 / 96.0 GB14–21 tok/s128KUnified memory
Runs wellDeepSeek-R1-Distill-Llama-8B DeepSeek8BQ8_09.4 / 96.0 GB14–21 tok/s125KUnified memory
Runs wellGemma 4 E4B Google8BQ8_08.6 / 96.0 GB16–24 tok/s125KUnified memory
Runs wellLing-3.0-tiny Ant Group7.9B / 1.3BBF16 (native)maker’s file15.6 / 96.0 GB48–72 tok/s128KUnified memory
Runs wellQwen2.5-7B-Instruct-1M Alibaba Qwen7.6BQ8_08.5 / 96.0 GB16–24 tok/s977KUnified memory
Runs wellQwen2.5-7B-Instruct Alibaba Qwen7.6BQ8_0maker’s file8.5 / 96.0 GB16–24 tok/s32KUnified memory
Runs wellDeepSeek-R1-Distill-Qwen-7B DeepSeek7.6BQ8_08.5 / 96.0 GB16–24 tok/s32KUnified memory
Runs wellMistral-7B-Instruct-v0.1 Mistral AI7.2BQ8_08.7 / 96.0 GB16–23 tok/s32KUnified memory
Runs wellGemma 4 E2B Google5.1BBF16 (native)maker’s file10.1 / 96.0 GB13–20 tok/s125KUnified memory
Runs wellWeVisOne-VL-4B Tencent4.8BBF16 (native)maker’s file10.6 / 96.0 GB13–19 tok/s256KUnified memory
Runs wellQianfan-OCR Baidu4.7BBF16 (native)maker’s file10.5 / 96.0 GB13–19 tok/s31KUnified memory
Runs wellQwen3.5-4B Alibaba Qwen4.7BBF16 (native)maker’s file9.4 / 96.0 GB14–21 tok/s256KUnified memory
Runs wellQwen3-VL-4B-Instruct Alibaba Qwen4.4BBF16 (native)maker’s file9.9 / 96.0 GB14–20 tok/s250KUnified memory
Runs wellQwen3-4B-Instruct-2507 Alibaba Qwen4BBF16 (native)maker’s file9.1 / 96.0 GB15–22 tok/s256KUnified memory
Runs wellQwen3-4B Alibaba Qwen4BBF16 (native)maker’s file9.1 / 96.0 GB15–22 tok/s128KUnified memory
Runs wellDeepSeek-OCR DeepSeek3.3BBF16 (native)maker’s file7.2 / 96.0 GB49–74 tok/s8KUnified memory
Runs wellGranite-4.0-H-Micro IBM3.2BBF16 (native)maker’s file6.5 / 96.0 GB21–32 tok/s128KUnified memory
Runs wellQwen3-1.7B Alibaba Qwen2BBF16 (native)maker’s file5.2 / 96.0 GB27–41 tok/s40KUnified memory
Runs wellDeepSeek-R1-Distill-Qwen-1.5B DeepSeek1.8BBF16 (native)maker’s file4.0 / 96.0 GB36–54 tok/s125KUnified memory
Runs wellGLM-OCR Zhipu AI1.3BBF16 (native)maker’s file3.5 / 96.0 GB43–64 tok/s128KUnified memory
Runs wellQwen3-0.6B Alibaba Qwen0.8BBF16 (native)maker’s file2.8 / 96.0 GB56–84 tok/s40KUnified memory
Runs wellEmbeddingGemma 2 Google0.7BBF16 (native)maker’s file2.0 / 96.0 GB87–130 tok/s256KUnified memory
Runs wellQwen3-Embedding-0.6B Alibaba Qwen0.6BBF16 (native)maker’s file2.5 / 96.0 GB64–96 tok/s32KUnified memory
Runs wellbge-reranker-v2-m3 BAAI0.6BFP32 (native)maker’s file3.4 / 96.0 GB44–67 tok/s8KUnified memory
Runs wellbge-m3-unsupervised BAAI0.6BFP32 (native)maker’s file3.4 / 96.0 GB44–67 tok/s8KUnified memory
Runs wellbge-large-en-v1.5 BAAI0.3BFP32 (native)maker’s file1.8 / 96.0 GB98–147 tok/s512Unified memory
Runs wellbge-reranker-base BAAI0.3BFP32 (native)maker’s file1.6 / 96.0 GB121–181 tok/s514Unified memory
Runs wellbge-base-en-v1.5 BAAI0.1BFP32 (native)maker’s file0.9 / 96.0 GB299–448 tok/s512Unified memory
Runs wellbge-small-en-v1.5 BAAI0BFP32 (native)maker’s file0.6 / 96.0 GB955–1433 tok/s512Unified memory
Runs, slowlyQwen2.5-VL-72B-Instruct Alibaba Qwen73.4BQ8_078.8 / 96.0 GB2–3 tok/s63KUnified memory
Runs, slowlyQwen2-VL-72B-Instruct Alibaba Qwen73.4BQ8_078.8 / 96.0 GB2–3 tok/s32KUnified memory
Runs, slowlyQwen2.5-72B-Instruct Alibaba Qwen72.7BQ8_0maker’s file78.3 / 96.0 GB2–3 tok/s32KUnified memory
Runs, slowlyQianfan-VL-70B Baidu71.8BQ8_077.1 / 96.0 GB2–3 tok/s32KUnified memory
Runs, slowlyDeepSeek-R1-Distill-Llama-70B DeepSeek70.6BQ8_075.8 / 96.0 GB2–3 tok/s72KUnified memory
Runs, slowlyQwen2.5-VL-32B-Instruct Alibaba Qwen33.5BBF16 (native)maker’s file68.8 / 96.0 GB2–3 tok/s117KUnified memory
Runs, slowlyQwen3-VL-32B-Instruct Alibaba Qwen33.4BBF16 (native)maker’s file67.2 / 96.0 GB2–3 tok/s123KUnified memory
Runs, slowlyQwen3-VL-32B-Thinking Alibaba Qwen33.4BBF16 (native)maker’s file67.2 / 96.0 GB2–3 tok/s123KUnified memory
Runs, slowlyQwen2.5-32B-Instruct Alibaba Qwen32.8BBF16 (native)maker’s file66.1 / 96.0 GB2–3 tok/s32KUnified memory
Runs, slowlyQwen2.5-Coder-32B-Instruct Alibaba Qwen32.8BBF16 (native)maker’s file66.1 / 96.0 GB2–3 tok/s32KUnified memory
Runs, slowlyQwQ-32B Alibaba Qwen32.8BBF16 (native)maker’s file66.1 / 96.0 GB2–3 tok/s40KUnified memory
Runs, slowlyDeepSeek-R1-Distill-Qwen-32B DeepSeek32.8BBF16 (native)maker’s file66.1 / 96.0 GB2–3 tok/s32KUnified memory
Runs, slowlyQwen3-32B Alibaba Qwen32.8BBF16 (native)maker’s file66.1 / 96.0 GB2–3 tok/s128KUnified memory
Runs, slowlyGemma 4 31B IT Google31.3BBF16 (native)maker’s file63.2 / 96.0 GB2–3 tok/s218KUnified memory
Runs, slowlyGranite Switch 4.1 30B IBM30BBF16 (native)maker’s file60.7 / 96.0 GB2–3 tok/s128KUnified memory
Runs, slowlyQwen3.5-27B Alibaba Qwen27.8BBF16 (native)maker’s file54.8 / 96.0 GB2–4 tok/s256KUnified memory
Runs, slowlyQwen3.6-27B Alibaba Qwen27.8BBF16 (native)maker’s file54.8 / 96.0 GB2–4 tok/s256KUnified memory
Runs, slowlyQwen3.8-27B Alibaba Qwen27.8BBF16 (native)maker’s file54.8 / 96.0 GB2–4 tok/s256KUnified memory
Runs, slowlyMagistral Small 1.2 Mistral AI24BQ8_0maker’s file25.8 / 96.0 GB5–8 tok/s125KUnified memory
Runs, slowlyDevstral Small 1.0 Mistral AI23.6BQ8_025.7 / 96.0 GB5–8 tok/s125KUnified memory
Runs, slowlyDevstral Small 1.1 Mistral AI23.6BQ8_025.7 / 96.0 GB5–8 tok/s125KUnified memory
Runs, slowlyMagistral Small 1.0 Mistral AI23.6BQ8_025.7 / 96.0 GB5–8 tok/s39KUnified memory
Runs, slowlyMagistral Small 1.1 Mistral AI23.6BQ8_0maker’s file25.8 / 96.0 GB5–8 tok/s39KUnified memory

Estimates from the makers’ Hugging Face repositories (as of 2026-11-06); filtered views aren’t indexed. How it’s estimated