← All machines

NVIDIA GB10 · NVIDIA

NVIDIA DGX Spark

NVIDIA's own GB10 desktop with 128GB of unified memory, CUDA and a ConnectX-7 port for linking up to four units. Its 273 GB/s memory bandwidth limits single-user token speed compared with Macs of similar memory size, while prompt processing is strong.

Memory for models

120 GBassumption

The maker publishes no GPU limit for this coherent memory pool; 8 GB is assumed for the operating system.

Memory bandwidth

273 GB/s

Maker price

~$4,699

Largest at 4-bit

Qwen3.8 Flash-Next

113.1 GB at 8,192 tokens of context (an assumed limit, see Memory for models)

gpt-oss-20b, 4-bit

83 tok/smedian of 13 runs

Memory configurations

One page covers every memory option the maker sells. Prices are the maker’s list price for the cheapest configuration with that much memory.

MemoryPriceMemory the GPU can use
128 GB~$4,699
120 GB gpu-usable memoryassumption

The maker publishes no GPU limit for this coherent memory pool; 8 GB is assumed for the operating system. How GPU memory limits work on unified-memory machines.

What it runs at 4-bit

Every model in the catalog at its 4-bit file with 8,192 tokens of context. 30 of 41 fit. Speeds are the median of measured runs on the same chip where any exist; otherwise an estimate from memory bandwidth. Try other quants and context lengths.

ModelNeedsFitGeneration speed
Kimi K32.8T, 104B active · UD-Q4_K_XL1406.8 GBcontext memory estimatedDoes not fit
DeepSeek V4 Pro1.6T, 49B active · Q4_K_M890.5 GBcontext memory not countedDoes not fit
Kimi K2.61T, 32B active · UD-Q4_K_XL545.7 GBcontext memory estimatedDoes not fit
DeepSeek V4.1 Flash552B, 16B active · Q4 (DwarfStar)484.5 GBDoes not fit
GLM-5.3753.3B · UD-Q4_K_XL437.4 GBcontext memory estimatedDoes not fit
DeepSeek V3.2671B, 37B active · Q4_K_M379.6 GBcontext memory estimatedDoes not fit
MiniMax M3428B, 23B active · Q4_K_M244.8 GBcontext memory not countedDoes not fit
Qwen3.5 397B-A17B397B, 17B active · Q4_K_M229.1 GBcontext memory estimatedDoes not fit
GLM-5.3 Flash320B, 18B active · UD-Q4_K_XL187.6 GBcontext memory estimatedDoes not fit
Hy3295B, 21B active · Q4_K_M173.6 GBDoes not fit
MiniMax M2.7228.7B · Q4_K_M132.7 GBDoes not fit
Qwen3.8 Flash-Next125B, 6B active · Q4_K_M113.1 GBcontext memory estimatedFitsassumed limit~22–29 tok/sestimate
Nemotron 3 Super 120B-A12B120B, 12B active · Q4_K_M82.6 GBcontext memory estimatedFitsassumed limit17 tok/smedian of 2 runs
Qwen3.5 122B-A10B122B, 10B active · Q4_K_M73.0 GBcontext memory estimatedFitsassumed limit21 tok/smedian of 2 runs
Mistral Medium 3.5 128B128B · Q4_K_M74.0 GBFitsassumed limit~2.5–3.0 tok/sestimate
GLM-4.5-Air106B, 12B active · Q4_K_M70.9 GBFitsassumed limit~16–20 tok/sestimate
Mistral Small 4 119B (2603)119B, 6.5B active · Q4_K_M69.3 GBcontext memory estimatedFitsassumed limit~32–41 tok/sestimate
Llama 4 Scout 17B-16E Instruct109B, 17B active · Q4_K_M62.4 GBcontext memory not countedFitsassumed limit~13–16 tok/sestimate
gpt-oss-120b117B, 5.1B active · MXFP4 (native)60.2 GBcontext memory estimatedFitsassumed limit53 tok/smedian of 23 runs
Qwen3 Coder Next80B, 3B active · Q4_K_M46.9 GBcontext memory estimatedFitsassumed limit49 tok/smedian of 2 runs
Llama 3.3 70B Instruct70B · Q4_K_M43.6 GBFitsassumed limit~4.4–5.2 tok/sestimate
Nemotron 3.5 Lightning 30B-A3B30B, 3B active · Q4_K_M25.3 GBcontext memory estimatedFitsassumed limit~50–64 tok/sestimate
Qwen3.6 35B-A3B35B, 3B active · Q4_K_M22.4 GBcontext memory estimatedFitsassumed limit~67–86 tok/sestimate
Ornith 1.5 35B-A3B35B, 3B active · Q4_K_M21.9 GBcontext memory estimatedFitsassumed limit~69–88 tok/sestimate
Qwen3 Coder 30B-A3B Instruct30.5B, 3.3B active · Q4_K_M19.5 GBFitsassumed limit~64–82 tok/sestimate
Gemma 4 31B IT30.7B · Q4_K_M18.9 GBcontext memory estimatedFitsassumed limit~10–12 tok/sestimate
Granite 4.2 30B29.3B · Q4_K_M20.3 GBFitsassumed limit~10–12 tok/sestimate
Qwen3.8 27B27B · Q4_K_M18.2 GBcontext memory estimatedFitsassumed limit~11–13 tok/sestimate
Gemma 4 26B-A4B IT25.2B, 3.8B active · Q4_K_M17.4 GBcontext memory estimatedFitsassumed limit~50–64 tok/sestimate
Qwen3.6 27B27B · Q4_K_M17.7 GBcontext memory estimatedFitsassumed limit~11–13 tok/sestimate
Devstral Small 2 24B (2512)24B · Q4_K_M16.1 GBFitsassumed limit~13–15 tok/sestimate
gpt-oss-20b21B, 3.6B active · MXFP4 (native)12.5 GBcontext memory estimatedFitsassumed limit83 tok/smedian of 13 runs
Phi-414.7B · Q4_K_M11.3 GBFitsassumed limit~21–25 tok/sestimate
Ministral 3 14B Reasoning (2512)13.9B · Q4_K_M10.4 GBFitsassumed limit~22–27 tok/sestimate
Gemma 4 12B IT11.95B · Q4_K_M8.2 GBcontext memory estimatedFitsassumed limit~26–31 tok/sestimate
Qwen3.5 9B9B · Q4_K_M7.0 GBcontext memory estimatedFitsassumed limit~33–39 tok/sestimate
Granite 4.2 8B8.8B · Q4_K_M7.9 GBFitsassumed limit~33–40 tok/sestimate
Gemma 4 E4B IT8B · Q4_K_M6.5 GBcontext memory not countedFitsassumed limit~61–73 tok/sestimate
Qwen3 8B8.2B · Q4_K_M7.3 GBFitsassumed limit~37–44 tok/sestimate
Llama 3.1 8B Instruct8B · Q4_K_M7.1 GBFitsassumed limit38 tok/smedian of 5 runs
Qwen3.5 4B4B · Q4_K_M4.3 GBcontext memory estimatedFitsassumed limit~68–81 tok/sestimate

Measured speeds

53 published runs with one user, plain decoding and at most 4,096 tokens already in context, on GB10 Grace Blackwell. Runs on Dell Pro Precision with GB10 with the same chip count here too, and say so.

Plain single-user runs
ModelQuantBackendGenerationPromptContextMeasured onSource
gpt-oss-120bQ4_K - Mediumllama.cpp b8580 CUDA62 tok/s2,444 tok/s2,048-token prompt0this machinellama.cpp benchmark discussion2026-03
gpt-oss-120bQ4_K - Mediumllama.cpp b8580 CUDA59 tok/s2,315 tok/s2,048-token prompt4,096this machinellama.cpp benchmark discussion2026-03
gpt-oss-120bMXFP4 MoEllama.cpp b7941 CUDA59 tok/s2,444 tok/s2,048-token prompt0this machinellama.cpp benchmark discussion2026-02
gpt-oss-120bMXFP4 MoEllama.cpp b7535 CUDA58 tok/s2,438 tok/s2,048-token prompt0this machinellama.cpp benchmark discussion2025-12
gpt-oss-120bMXFP4 MoEllama.cpp (build not stated) CUDA57 tok/s1,816 tok/s2,048-token prompt0a machine with the same chipllama.cpp benchmark discussion2025-11
gpt-oss-120bMXFP4 MoEllama.cpp b6817 CUDA56 tok/s1,939 tok/s2,048-token prompt0this machinellama.cpp benchmark discussion2025-10
gpt-oss-120bMXFP4llama.cpp b707056 tok/s966 tok/s2,048-token prompt0Dell Pro Precision with GB10 (same chip)owner report2025-12
gpt-oss-120bMXFP4 MoEllama.cpp b7941 CUDA56 tok/s2,310 tok/s2,048-token prompt4,096this machinellama.cpp benchmark discussion2026-02
gpt-oss-120bMXFP4llama.cpp (build not stated)55 tok/s1,725 tok/s2,048-token prompt0this machinemaker2025-10
gpt-oss-120bMXFP4 MoEllama.cpp b6817 CUDA55 tok/s1,738 tok/s2,048-token prompt0this machinellama.cpp benchmark discussion2025-10
gpt-oss-120bMXFP4 MoEllama.cpp b7535 CUDA55 tok/s2,294 tok/s2,048-token prompt4,096this machinellama.cpp benchmark discussion2025-12
gpt-oss-120bMXFP4 MoEllama.cpp b6817 CUDA53 tok/s1,832 tok/s2,048-token prompt4,096this machinellama.cpp benchmark discussion2025-10
gpt-oss-120bMXFP4llama.cpp (build not stated)53 tok/s1,729 tok/s2,048-token prompt0a machine with the same chipowner report2025-12
gpt-oss-120bMXFP4 MoEllama.cpp b6817 CUDA51 tok/s1,778 tok/s2,048-token prompt4,096this machinellama.cpp benchmark discussion2025-10
gpt-oss-120bMXFP4SGLang (lmsysorg/sglang:spark image)51 tok/s676 tok/s2,048-token prompt0this machinereview2025-10
gpt-oss-120bMXFP4llama.cpp (build not stated)49 tok/s1,608 tok/s2,048-token prompt0this machineowner report2025-12
gpt-oss-120bMXFP4 MoEllama.cpp b6816 CUDA46 tok/s1,737 tok/s2,048-token prompt0this machinellama.cpp benchmark discussion2025-10
gpt-oss-120bMXFP4 MoEllama.cpp b8322 CUDA45 tok/s2,047 tok/s4,096-token prompt0this machinellama.cpp benchmark discussion2026-03
gpt-oss-120bMXFP4 MoEllama.cpp b6816 CUDA43 tok/s1,778 tok/s2,048-token prompt4,096this machinellama.cpp benchmark discussion2025-10
gpt-oss-120bMXFP4Ollama (version not stated)42 tok/s95 tok/s0this machinereview2025-10
gpt-oss-120bMXFP4 MoEllama.cpp (build not stated) CUDA42 tok/s1,864 tok/s2,048-token prompt0this machinellama.cpp benchmark discussion2025-10
gpt-oss-120bMXFP4Ollama v0.12.641 tok/s1,169 tok/s0this machineOllama blog2025-10
gpt-oss-120bMXFP4 MoEllama.cpp (build not stated) CUDA38 tok/s1,731 tok/s2,048-token prompt4,096this machinellama.cpp benchmark discussion2025-10
gpt-oss-20bMXFP4 MoEllama.cpp b7030 Metal,BLAS118 tok/s–0this machinellama.cpp benchmark discussion2025-11
gpt-oss-20bMXFP4 MoEllama.cpp b7067 CUDA86 tok/s3,798 tok/s2,048-token prompt0this machinellama.cpp benchmark discussion2025-11
gpt-oss-20bMXFP4 MoEllama.cpp b7030 CUDA86 tok/s–0this machinellama.cpp benchmark discussion2025-11
gpt-oss-20bQ4_K - Mediumllama.cpp b8580 CUDA86 tok/s4,453 tok/s2,048-token prompt0this machinellama.cpp benchmark discussion2026-03
gpt-oss-20bMXFP4 MoEllama.cpp (build not stated) CUDA84 tok/s3,730 tok/s2,048-token prompt0a machine with the same chipllama.cpp benchmark discussion2025-11
gpt-oss-20bMXFP4 MoEllama.cpp b7941 CUDA83 tok/s4,506 tok/s2,048-token prompt0this machinellama.cpp benchmark discussion2026-02
gpt-oss-20bMXFP4llama.cpp (build not stated)83 tok/s3,670 tok/s2,048-token prompt0this machinemaker2025-10
gpt-oss-20bQ4_K - Mediumllama.cpp b8580 CUDA82 tok/s4,230 tok/s2,048-token prompt4,096this machinellama.cpp benchmark discussion2026-03
gpt-oss-20bMXFP4 MoEllama.cpp b7941 CUDA79 tok/s4,158 tok/s2,048-token prompt4,096this machinellama.cpp benchmark discussion2026-02
gpt-oss-20bMXFP4llama.cpp b707079 tok/s1,646 tok/s2,048-token prompt0Dell Pro Precision with GB10 (same chip)owner report2025-12
gpt-oss-20bMXFP4SGLang (lmsysorg/sglang:spark image)70 tok/s434 tok/s2,048-token prompt0this machinereview2025-10
gpt-oss-20bMXFP4Ollama (version not stated)61 tok/s2,054 tok/s0this machinereview2025-10
gpt-oss-20bMXFP4Ollama v0.12.658 tok/s3,224 tok/s0this machineOllama blog2025-10
Llama 3.1 8B InstructQ4_K_MOllama (version not stated)43 tok/s23,170 tok/s0this machinereview2025-10
Llama 3.1 8B InstructNVFP4TensorRT-LLM (version not stated)39 tok/s10,257 tok/s2,048-token prompt0this machinemaker2025-10
Llama 3.1 8B InstructQ4_K_MOllama v0.12.638 tok/s7,614 tok/s0this machineOllama blog2025-10
Llama 3.1 8B InstructQ4_K_MLocalScore (llamafile; version not stated)35 tok/s2,010 tok/s0a machine with the same chipLocalScore2026-10
Llama 3.1 8B InstructQ4_K_MLocalScore (llamafile; version not stated)34 tok/s1,986 tok/s0a machine with the same chipLocalScore2026-10
Llama 3.1 8B InstructQ8_0Ollama (version not stated)29 tok/s19,826 tok/s0this machinereview2025-10
Llama 3.1 8B InstructQ8_0Ollama v0.12.625 tok/s6,110 tok/s0this machineOllama blog2025-10
Llama 3.1 8B InstructFP8SGLang (lmsysorg/sglang:spark image)21 tok/s7,991 tok/s2,048-token prompt0this machinereview2025-10
Nemotron 3 Super 120B-A12BMXFP4 MoEllama.cpp b8580 CUDA17 tok/s730 tok/s2,048-token prompt0this machinellama.cpp benchmark discussion2026-03
Nemotron 3 Super 120B-A12BMXFP4 MoEllama.cpp b8580 CUDA17 tok/s725 tok/s2,048-token prompt4,096this machinellama.cpp benchmark discussion2026-03
Qwen3 Coder 30B-A3B InstructQ8_0llama.cpp b7941 CUDA61 tok/s2,987 tok/s2,048-token prompt0this machinellama.cpp benchmark discussion2026-02
Qwen3 Coder 30B-A3B InstructQ8_0llama.cpp b7941 CUDA55 tok/s2,633 tok/s2,048-token prompt4,096this machinellama.cpp benchmark discussion2026-02
Qwen3 Coder 30B-A3B InstructQ8_0llama.cpp (build not stated)55 tok/s2,705 tok/s2,048-token prompt0this machineowner report2025-12
Qwen3 Coder NextMXFP4 MoEllama.cpp b8580 CUDA49 tok/s1,743 tok/s2,048-token prompt0this machinellama.cpp benchmark discussion2026-03
Qwen3 Coder NextMXFP4 MoEllama.cpp b8580 CUDA49 tok/s1,708 tok/s2,048-token prompt4,096this machinellama.cpp benchmark discussion2026-03
Qwen3.5 122B-A10BMXFP4_MOE (unsloth)llama.cpp b8580 CUDA21 tok/s1,075 tok/s2,048-token prompt0this machinellama.cpp benchmark discussion2026-03
Qwen3.5 122B-A10BMXFP4_MOE (unsloth)llama.cpp b8580 CUDA21 tok/s1,050 tok/s2,048-token prompt4,096this machinellama.cpp benchmark discussion2026-03

With speculative decoding

A small draft model or extra prediction heads guess several tokens ahead. Generation is faster on text the guesses get right, so these are kept out of the medians above.

Runs with speculative decoding
ModelQuantBackendGenerationPromptContextMeasured onSource
Qwen3.8 27BQ4_K_Mllama.cpp (ghcr.io full-cuda Docker image, build not stated)18 tok/s489 tok/s6,231-token prompt6,231this machinellama.cpp benchmark discussion2026-08
Qwen3.8 27BUD-Q4_K_XLllama.cpp (ghcr.io full-cuda Docker image, build not stated)15 tok/s––this machinellama.cpp benchmark discussion2026-08

Many users at once

Total throughput with several requests in flight. One person chatting sees far less than this.

Runs with several requests at once
ModelQuantBackendGenerationPromptContextMeasured onSource
gpt-oss-120bMXFP4SGLang (lmsysorg/sglang:spark image)32 requests at once203 tok/s307 tok/s2,048-token prompt0this machinereview2025-10
gpt-oss-120bMXFP4SGLang (lmsysorg/sglang:spark image)8 requests at once117 tok/s302 tok/s2,048-token prompt0this machinereview2025-10
gpt-oss-20bMXFP4SGLang (lmsysorg/sglang:spark image)32 requests at once522 tok/s546 tok/s2,048-token prompt0this machinereview2025-10
gpt-oss-20bMXFP4SGLang (lmsysorg/sglang:spark image)8 requests at once219 tok/s553 tok/s2,048-token prompt0this machinereview2025-10
Llama 3.1 8B InstructFP8SGLang (lmsysorg/sglang:spark image)32 requests at once368 tok/s7,950 tok/s2,048-token prompt0this machinereview2025-10
Llama 3.1 8B InstructFP8SGLang (lmsysorg/sglang:spark image)8 requests at once144 tok/s7,744 tok/s2,048-token prompt0this machinereview2025-10

Deep into a long context

Measured with more than 4,096 tokens already in context, where every new token also reads the context memory.

Runs at long context
ModelQuantBackendGenerationPromptContextMeasured onSource
gpt-oss-120bQ4_K - Mediumllama.cpp b8580 CUDA56 tok/s2,222 tok/s2,048-token prompt8,192this machinellama.cpp benchmark discussion2026-03
gpt-oss-120bMXFP4 MoEllama.cpp b7941 CUDA53 tok/s2,217 tok/s2,048-token prompt8,192this machinellama.cpp benchmark discussion2026-02
gpt-oss-120bMXFP4 MoEllama.cpp b7535 CUDA52 tok/s2,149 tok/s2,048-token prompt8,192this machinellama.cpp benchmark discussion2025-12
gpt-oss-120bQ4_K - Mediumllama.cpp b8580 CUDA51 tok/s1,948 tok/s2,048-token prompt16,384this machinellama.cpp benchmark discussion2026-03
gpt-oss-120bMXFP4 MoEllama.cpp b7941 CUDA49 tok/s1,956 tok/s2,048-token prompt16,384this machinellama.cpp benchmark discussion2026-02
gpt-oss-120bMXFP4 MoEllama.cpp b6817 CUDA49 tok/s1,738 tok/s2,048-token prompt8,192this machinellama.cpp benchmark discussion2025-10
gpt-oss-120bMXFP4 MoEllama.cpp b7535 CUDA48 tok/s1,824 tok/s2,048-token prompt16,384this machinellama.cpp benchmark discussion2025-12
gpt-oss-120bMXFP4 MoEllama.cpp b6817 CUDA48 tok/s1,714 tok/s2,048-token prompt8,192this machinellama.cpp benchmark discussion2025-10
gpt-oss-120bMXFP4 MoEllama.cpp b6817 CUDA45 tok/s1,526 tok/s2,048-token prompt16,384this machinellama.cpp benchmark discussion2025-10
gpt-oss-120bMXFP4 MoEllama.cpp b6817 CUDA44 tok/s1,533 tok/s2,048-token prompt16,384this machinellama.cpp benchmark discussion2025-10
gpt-oss-120bQ4_K - Mediumllama.cpp b8580 CUDA44 tok/s1,561 tok/s2,048-token prompt32,768this machinellama.cpp benchmark discussion2026-03
gpt-oss-120bMXFP4llama.cpp (build not stated)43 tok/s1,357 tok/s2,048-token prompt16,384a machine with the same chipowner report2025-12
gpt-oss-120bMXFP4 MoEllama.cpp b7941 CUDA43 tok/s1,567 tok/s2,048-token prompt32,768this machinellama.cpp benchmark discussion2026-02
gpt-oss-120bMXFP4llama.cpp b707042 tok/s305 tok/s2,048-token prompt16,384Dell Pro Precision with GB10 (same chip)owner report2025-12
gpt-oss-120bMXFP4 MoEllama.cpp b6816 CUDA42 tok/s1,720 tok/s2,048-token prompt8,192this machinellama.cpp benchmark discussion2025-10
gpt-oss-120bMXFP4 MoEllama.cpp b7535 CUDA41 tok/s1,416 tok/s2,048-token prompt32,768this machinellama.cpp benchmark discussion2025-12
gpt-oss-120bMXFP4llama.cpp (build not stated)40 tok/s1,308 tok/s2,048-token prompt16,384this machineowner report2025-12
gpt-oss-120bMXFP4 MoEllama.cpp b6817 CUDA39 tok/s1,242 tok/s2,048-token prompt32,768this machinellama.cpp benchmark discussion2025-10
gpt-oss-120bMXFP4 MoEllama.cpp b6817 CUDA39 tok/s1,238 tok/s2,048-token prompt32,768this machinellama.cpp benchmark discussion2025-10
gpt-oss-120bMXFP4 MoEllama.cpp b6816 CUDA38 tok/s1,512 tok/s2,048-token prompt16,384this machinellama.cpp benchmark discussion2025-10
gpt-oss-120bMXFP4llama.cpp (build not stated)37 tok/s1,074 tok/s2,048-token prompt32,768a machine with the same chipowner report2025-12
gpt-oss-120bMXFP4 MoEllama.cpp b8322 CUDA37 tok/s1,676 tok/s4,096-token prompt20,000this machinellama.cpp benchmark discussion2026-03
gpt-oss-120bMXFP4 MoEllama.cpp (build not stated) CUDA36 tok/s1,628 tok/s2,048-token prompt8,192this machinellama.cpp benchmark discussion2025-10
gpt-oss-120bMXFP4llama.cpp (build not stated)34 tok/s1,061 tok/s2,048-token prompt32,768this machineowner report2025-12
gpt-oss-120bMXFP4 MoEllama.cpp b6816 CUDA34 tok/s1,232 tok/s2,048-token prompt32,768this machinellama.cpp benchmark discussion2025-10
gpt-oss-120bMXFP4 MoEllama.cpp (build not stated) CUDA34 tok/s1,395 tok/s2,048-token prompt16,384this machinellama.cpp benchmark discussion2025-10
gpt-oss-120bMXFP4llama.cpp b707033 tok/s181 tok/s2,048-token prompt32,768Dell Pro Precision with GB10 (same chip)owner report2025-12
gpt-oss-120bMXFP4 MoEllama.cpp b8322 CUDA30 tok/s1,225 tok/s4,096-token prompt48,000this machinellama.cpp benchmark discussion2026-03
gpt-oss-20bQ4_K - Mediumllama.cpp b8580 CUDA77 tok/s3,982 tok/s2,048-token prompt8,192this machinellama.cpp benchmark discussion2026-03
gpt-oss-20bMXFP4 MoEllama.cpp b7941 CUDA75 tok/s3,994 tok/s2,048-token prompt8,192this machinellama.cpp benchmark discussion2026-02
gpt-oss-20bQ4_K - Mediumllama.cpp b8580 CUDA72 tok/s3,366 tok/s2,048-token prompt16,384this machinellama.cpp benchmark discussion2026-03
gpt-oss-20bMXFP4 MoEllama.cpp b7941 CUDA70 tok/s3,450 tok/s2,048-token prompt16,384this machinellama.cpp benchmark discussion2026-02
gpt-oss-20bQ4_K - Mediumllama.cpp b8580 CUDA62 tok/s2,616 tok/s2,048-token prompt32,768this machinellama.cpp benchmark discussion2026-03
gpt-oss-20bMXFP4 MoEllama.cpp b7941 CUDA62 tok/s2,689 tok/s2,048-token prompt32,768this machinellama.cpp benchmark discussion2026-02
gpt-oss-20bMXFP4llama.cpp b707061 tok/s474 tok/s2,048-token prompt16,384Dell Pro Precision with GB10 (same chip)owner report2025-12
gpt-oss-20bMXFP4llama.cpp b707048 tok/s271 tok/s2,048-token prompt32,768Dell Pro Precision with GB10 (same chip)owner report2025-12
Nemotron 3 Super 120B-A12BMXFP4 MoEllama.cpp b8580 CUDA17 tok/s721 tok/s2,048-token prompt8,192this machinellama.cpp benchmark discussion2026-03
Nemotron 3 Super 120B-A12BMXFP4 MoEllama.cpp b8580 CUDA16 tok/s714 tok/s2,048-token prompt16,384this machinellama.cpp benchmark discussion2026-03
Nemotron 3 Super 120B-A12BMXFP4 MoEllama.cpp b8580 CUDA16 tok/s696 tok/s2,048-token prompt32,768this machinellama.cpp benchmark discussion2026-03
Qwen3 Coder 30B-A3B InstructQ8_0llama.cpp b7941 CUDA48 tok/s2,354 tok/s2,048-token prompt8,192this machinellama.cpp benchmark discussion2026-02
Qwen3 Coder 30B-A3B InstructQ8_0llama.cpp b7941 CUDA40 tok/s1,909 tok/s2,048-token prompt16,384this machinellama.cpp benchmark discussion2026-02
Qwen3 Coder 30B-A3B InstructQ8_0llama.cpp (build not stated)35 tok/s1,631 tok/s2,048-token prompt16,384this machineowner report2025-12
Qwen3 Coder 30B-A3B InstructQ8_0llama.cpp b7941 CUDA30 tok/s1,348 tok/s2,048-token prompt32,768this machinellama.cpp benchmark discussion2026-02
Qwen3 Coder 30B-A3B InstructQ8_0llama.cpp (build not stated)26 tok/s1,169 tok/s2,048-token prompt32,768this machineowner report2025-12
Qwen3 Coder NextMXFP4 MoEllama.cpp b8580 CUDA47 tok/s1,678 tok/s2,048-token prompt8,192this machinellama.cpp benchmark discussion2026-03
Qwen3 Coder NextMXFP4 MoEllama.cpp b8580 CUDA45 tok/s1,607 tok/s2,048-token prompt16,384this machinellama.cpp benchmark discussion2026-03
Qwen3 Coder NextMXFP4 MoEllama.cpp b8580 CUDA41 tok/s1,493 tok/s2,048-token prompt32,768this machinellama.cpp benchmark discussion2026-03
Qwen3.5 122B-A10BMXFP4_MOE (unsloth)llama.cpp b8580 CUDA21 tok/s1,026 tok/s2,048-token prompt8,192this machinellama.cpp benchmark discussion2026-03
Qwen3.5 122B-A10BMXFP4_MOE (unsloth)llama.cpp b8580 CUDA20 tok/s976 tok/s2,048-token prompt16,384this machinellama.cpp benchmark discussion2026-03
Qwen3.5 122B-A10BMXFP4_MOE (unsloth)llama.cpp b8580 CUDA20 tok/s897 tok/s2,048-token prompt32,768this machinellama.cpp benchmark discussion2026-03

Specifications

Chip
GB10 Grace Blackwell
Memory
LPDDR5X, 256-bit, coherent unified
Memory bandwidth
273 GB/s
FP16 compute
NVIDIA quotes up to 1 PFLOP FP4 with sparsity only; no dense FP16 figure
Software
CUDA, Vulkan
Ships with
NVIDIA DGX OS (Ubuntu-based)
Power
240 W (240W power supply (NVIDIA spec); GB10 TDP 140W)
Form factor
Mini PC
Networking
10GbE RJ-45; ConnectX-7 at 200Gbps (two QSFP ports): two or three units cluster by direct cable, four need a 200GbE switch
Released
2025-10
Maker page
nvidia.com

Sources

  • nvidia.com: GB10 specs: 20-core Arm CPU, 128GB LPDDR5x 256-bit, 273 GB/s, 1 PFLOP FP4 (sparse), 10GbE + ConnectX-7 200Gbps, 240W PSU, 140W GB10 TDP, DGX OS, cluster up to 4; 64GB config only via OEMs
  • owner discussion: MSRP changed from $3,999 to $4,699
  • nvidianews.nvidia.com: orderable from Oct 15, 2025; OEM partner list
  • openclawdc.com: Founders Edition reported sold out early September 2026 (stock context only)

Compared with

Similar machines