DGX Spark · 128 GB unified
NVIDIA · DGX Spark
193 text models run well · 273 GB/s bandwidth
Models on this hardware
Start here
File UD-Q4_K_M · 16.9 GB · unsloth/gemma-4-26B-A4B-it-GGUF · What is Q4?
Memory Needs ~19 GB · ~27 GB on DGX Spark · 128 GB unified, including ~8 GB the system keeps
ollama run hf.co/unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_M- Install Ollama
- Run the command above
- Ask:
Explain what a context window is, in two sentences.
File Q4_K_M · 16.5 GB · prithivMLmods/Smaug-Mini-GGUF · What is Q4?
Memory Needs ~21 GB · ~29 GB on DGX Spark · 128 GB unified, including ~8 GB the system keeps
ollama run hf.co/prithivMLmods/Smaug-Mini-GGUF:Q4_K_M- Install Ollama
- Run the command above
- Ask:
Explain what a context window is, in two sentences.
Text
Text · LiveBench · overall
0–100 · higher is better
- Color = publisher
Source: LiveBench 2026-06-25 · collected 2026-09-26 · Open models without a score are not drawn (no score ≠ 0).
| Smaug MiniQ4 · published params · source | 76.9 | Not on the LMArena text leaderboard of 2026-10-22 | ~29 GB | Q4 · published params |
|---|---|---|---|---|
| Qwen3.8 27BQ4 · published params · source | 75.3 | 1,437 ± 5 | ~29 GB | Q4 · published params |
| Qwen3.6 27BQ4 · published params · source | 64.0 | Not on the LMArena text leaderboard of 2026-10-22 | ~29 GB | Q4 · published params |
| Qwen3 14BQ4 · published params · source | No comparable published result | Not on the LMArena text leaderboard of 2026-10-22 | ~20 GB | Q4 · published params |
| Gemma 4 31BQ4 · published params · source | No comparable published result | 1,452 ± 7 | ~32 GB | Q4 · published params |
| Gemma 4 26B A4BQ4 · published params · source | No comparable published result | 1,437 ± 7 | ~27 GB | Q4 · published params |
| Qwen3 8BQ4 · published params · source | No comparable published result | Not on the LMArena text leaderboard of 2026-10-22 | ~15 GB | Q4 · published params |
| Qwen3.5 9BQ4 · published params · source | No comparable published result | Not on the LMArena text leaderboard of 2026-10-22 | ~15 GB | Q4 · published params |
LMArena ratings in this table: LMArena 2026-10-22 · collected 2026-10-25 · dataset · CC BY 4.0
Tight fit: it runs, but leaves less than 12.8 GB free beyond the 8 GB the system keeps (the memory figure already includes that reserve).
Every reviewed text model here 39
Every one that runs here, well or slowly, scored or not; 19 of 39 have a LiveBench or LMArena result. Run well first, then tight fits, then slow; within each, by Hub downloads over 30 days, read 2026-10-25, not by quality.
| Qwen3 14B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~20 GB | 9.0 | Apache 2.0 | 10.4M |
|---|---|---|---|---|---|---|
| Gemma 4 31B | Not in LiveBench 2026-06-25 | 1,452 | ~32 GB | 18.3 | Apache 2.0 | 8.4M |
| Gemma 4 26B A4B | Not in LiveBench 2026-06-25 | 1,437 | ~27 GB | 16.9 | Apache 2.0 | 8.2M |
| Qwen3 8B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~15 GB | 5.0 | Apache 2.0 | 8.1M |
| Qwen3.8 27B | 75.3 | 1,437 | ~29 GB | 16.5 | Apache 2.0 | 7.8M |
| Qwen3.5 9B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~15 GB | 5.7 | Apache 2.0 | 6.1M |
| gpt-oss-20b | Not in LiveBench 2026-06-25 | 1,317 | ~24 GB | 11.6 | Apache 2.0 | 5.8M |
| Llama 3.1 8B Instruct | Not in LiveBench 2026-06-25 | 1,211 | ~14 GB | 4.9 | Llama 3.1 | 4.8M |
| Qwen3 1.7B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~10 GB | 1.1 | Apache 2.0 | 4.1M |
| Gemma 4 E4B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~14 GB | 5.0 | Apache 2.0 | 4M |
| Llama 3.2 1B Instruct | Not in LiveBench 2026-06-25 | 1,111 | ~9 GB | 0.8 | Llama 3.2 | 3.5M |
| Gemma 4 E2B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~12 GB | 3.1 | Apache 2.0 | 3M |
| Qwen3 4B Instruct 2507 | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~11 GB | 2.5 | Apache 2.0 | 2.6M |
| Qwen3.6 27B | 64.0 | Not on the LMArena text leaderboard of 2026-10-22 | ~29 GB | 16.8 | Apache 2.0 | 2.4M |
| Gemma 4 12B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~17 GB | 7.1 | Apache 2.0 | 1.7M |
| Qwen3.5 35B A3B | Not in LiveBench 2026-06-25 | 1,394 | ~35 GB | 22.0 | Apache 2.0 | 1.5M |
| Nemotron 3 Nano 30B A3B | Not in LiveBench 2026-06-25 | 1,314 | ~38 GB | 24.6 | NVIDIA Nemotron Open Model | 1.4M |
| Llama 3.2 3B Instruct | Not in LiveBench 2026-06-25 | 1,167 | ~11 GB | 2.0 | Llama 3.2 | 1.1M |
| Qwen3 30B A3B Instruct 2507 | Not in LiveBench 2026-06-25 | 1,382 | ~31 GB | 18.6 | Apache 2.0 | 1.1M |
| GLM-4.7 Flash | Not in LiveBench 2026-06-25 | 1,365 | ~31 GB | 18.3 | MIT | 953.7K |
| Nemotron 3.5 Lightning 30B A3B | Not in LiveBench 2026-06-25 | Arena rated the NVFP4 build, NVIDIA-Nemotron-3.5-Lightning-30B-A3B (NVFP4); this entry is the BF16 reference, which NVIDIA reports with different scores | ~31 GB | 18.9 | OpenMDW 1.1 | 861.5K |
| Qwen3 Coder 30B A3B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~31 GB | 18.6 | Apache 2.0 | 745K |
| SmolLM3 3B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~11 GB | 1.9 | Apache 2.0 | 718.7K |
| Nemotron 3 Super 120B A12B | Not in LiveBench 2026-06-25 | Arena does not state the precision of the build it served, so no rating is matched to this NVFP4 checkpoint | ~98 GB | 82.5 | NVIDIA Nemotron Open Model | 535.1K |
| Phi-4 | Not in LiveBench 2026-06-25 | 1,256 | ~19 GB | 8.0 | MIT | 368.3K |
| DeepSeek R1 0528 Qwen3 8B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~15 GB | 5.0 | MIT | 355.8K |
| Phi-4 mini | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~11 GB | 2.5 | MIT | 342.8K |
| Olmo 3 7B Instruct | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~14 GB | 4.5 | Apache 2.0 | 262.5K |
| Devstral Small 2 24B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~26 GB | 14.3 | Apache 2.0 | 195.6K |
| Mistral Small 3.2 24B | Not in LiveBench 2026-06-25 | 1,356 | ~27 GB | 14.3 | Apache 2.0 | 181.4K |
| Ministral 3 14B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~19 GB | 8.2 | Apache 2.0 | 126.5K |
| Ministral 3 3B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~11 GB | 2.1 | Apache 2.0 | 87.4K |
| Ministral 3 8B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~15 GB | 5.2 | Apache 2.0 | 79.4K |
| Granite 4.2 8B | Not in LiveBench 2026-06-25 | 1,290 | ~14 GB | 5.3 | Apache 2.0 | 78.3K |
| Granite 4.2 3B | Not in LiveBench 2026-06-25 | 1,290 | ~11 GB | 2.2 | Apache 2.0 | 39.8K |
| Granite 4.2 30B | Not in LiveBench 2026-06-25 | 1,340 | ~31 GB | 17.7 | Apache 2.0 | 39.3K |
| Olmo 3 32B Think | Not in LiveBench 2026-06-25 | 1,306 | ~32 GB | 19.5 | Apache 2.0 | 13.1K |
| Ling 3.0 Flash | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~101 GB | 77.8 | MIT | 9.7K |
| Smaug Mini | 76.9 | Not on the LMArena text leaderboard of 2026-10-22 | ~29 GB | 16.5 | Apache 2.0 | 817 |
With 256 GB you also run… 6
The next memory size on this ladder runs these well. Figures are the requirement on this hardware type.
- DeepSeek V4 FlashDeepSeek~221 GBQ4 · published params
- DeepSeek V4 Flash · 0731DeepSeek~236 GBQ4 · published params
- MiMo V2.6 FlashXiaomi~240 GBQ4 · published params
- GLM-5.3 FlashZ.ai~248 GBQ4 · published params
- Qwen3-235B-A22B-Instruct-2507Qwen~185 GBQ4 · tensor count
- Qwen3-235B-A22B-Thinking-2507Qwen~185 GBQ4 · tensor count
Adjust the estimate: quantization, context, reserve
Vision
Vision · LMArena
rating · higher is better
- Color = publisher
Source: LMArena 2026-10-19 · collected 2026-10-25 · dataset · CC BY 4.0 · Models without a rating are left out, not scored as zero.
| Gemma 4 31BQ4 · published params · source | 1,262 ± 6 | ~32 GB | Q4 · published params |
|---|---|---|---|
| Qwen3.8 27BQ4 · published params · source | 1,242 ± 8 | ~29 GB | Q4 · published params |
| Gemma 4 26B A4BQ4 · published params · source | 1,242 ± 6 | ~27 GB | Q4 · published params |
| gemma-3-27b-itQ4 · Hugging Face count · source | 1,160 ± 8 | ~29 GB | Q4 · Hugging Face count |
| GLM-4.5VQ4 · Hugging Face count · source | 1,153 ± 12 | ~89 GB | Q4 · Hugging Face count |
| Llama-4-Scout-17B-16E-InstructQ4 · Hugging Face count · source | 1,128 ± 10 | ~90 GB | Q4 · Hugging Face count |
| Qwen2.5-VL-32B-InstructQ4 · Hugging Face count · source | 1,119 ± 15 | ~34 GB | Q4 · Hugging Face count |
| Molmo2-8BQ4 · Hugging Face count · source | 1,112 ± 20 | ~15 GB | Q4 · Hugging Face count |
LMArena ratings in this table: LMArena 2026-10-19 · collected 2026-10-25 · dataset · CC BY 4.0
With 256 GB you also run… 4
The next memory size on this ladder runs these well. Figures are the requirement on this hardware type.
- GLM-5.3 FlashZ.ai~248 GBQ4 · published params
- Qwen3-VL-235B-A22B-InstructQwen~185 GBQ4 · Hugging Face count
- Intern-S1internlm~189 GBQ4 · Hugging Face count
- step3stepfun-ai~249 GBQ4 · Hugging Face count
12 more vision models in the catalog have no usable memory figure or parameter count for this profile. Browse Vision models
Image
Image · Prompt alignment · GenEval
0–1 · higher is better
Source: GenEval · 2023 · reviewed 2026-10-05 · Published quality in the original benchmark preset; not measured on this hardware.
| Stable Diffusion XL 1.0Exact hardware documented · TensorRT / NGC PyTorch · source | 0.55 | Not on the LMArena text-to-image leaderboard of 2026-10-07 | Documented path without a memory figure | Exact hardware documented |
|---|---|---|---|---|
| Stable Diffusion 1.516-bit est. · sum of parts · source | 0.43 | Not on the LMArena text-to-image leaderboard of 2026-10-07 | ~5 GB | 16-bit est. · sum of parts |
| Emu3.5-Image16-bit est. · params · source | No comparable published result | Not reviewed for LMArena | ~84 GB | 16-bit est. · params |
| Ming-Image-0.1-DesignConditional path · CUDA · source | No comparable published result | Not reviewed for LMArena | ~80 GB | Conditional path |
| Qwen-Image-Edit16-bit est. · params · source | No comparable published result | Not reviewed for LMArena | ~51 GB | 16-bit est. · params |
| Qwen-Image-Edit-250916-bit est. · params · source | No comparable published result | Not reviewed for LMArena | ~51 GB | 16-bit est. · params |
| Qwen-Image-Edit-251116-bit est. · params · source | No comparable published result | Not reviewed for LMArena | ~51 GB | 16-bit est. · params |
| Qwen-Image-Layered16-bit est. · params · source | No comparable published result | Not reviewed for LMArena | ~51 GB | 16-bit est. · params |
LMArena ratings in this table: LMArena 2026-10-07 · collected 2026-10-25 · dataset · CC BY 4.0
With 256 GB you also run… 1
The next memory size on this ladder. Figures are the requirement on this hardware type.
- HunyuanImage 3.0 (Fal)Tencent~240 GBPublisher minimum
17 more image models in the catalog have no usable memory figure or parameter count for this profile. Browse Image models
Video
| MiniMax H316-bit est. · params · source | ~82 GB | 16-bit est. · params |
|---|---|---|
| Cosmos-H-SurgicalConditional path · CUDA · source | ~80 GB | Conditional path |
| Wan-Dancer-14BPublisher minimum · source | ~80 GB | Publisher minimum |
| Wan2.2 T2V-A14BConditional path · CUDA · source | ~80 GB | Conditional path |
| Wan2.2-I2V-A14BConditional path · CUDA · source | ~80 GB | Conditional path |
| Wan2.2-I2V-A14B-DiffusersConditional path · CUDA · source | ~80 GB | Conditional path |
| Wan2.2-S2V-14BConditional path · CUDA · source | ~80 GB | Conditional path |
| Wan2.2-T2V-A14B-DiffusersConditional path · CUDA · source | ~80 GB | Conditional path |
How this list works Documented execution paths on this hardware first, then publisher minimums (CUDA or Apple MPS).
With 256 GB you also run… 2
The next memory size on this ladder. Figures are the requirement on this hardware type.
- Cosmos3-Super-Image2Video-4Stepnvidia~156 GB16-bit est. · params
- Cosmos3-Super-Image2Videonvidia~158 GB16-bit est. · params
20 more video models in the catalog have no usable memory figure or parameter count for this profile. Browse Video models
Audio
Audio · Word error rate · Open ASR
% · lower is better
Source: Word error rate · Open ASR · reviewed 2026-10-20 · Published quality in the original benchmark preset; not measured on this hardware.
| Whisper large-v3-turboConditional path · CUDA · source | 7.83 | ~6 GB | Conditional path |
|---|---|---|---|
| Whisper large-v3Conditional path · CUDA · source | 7.44 | ~10 GB | Conditional path |
| Step-Audio-R1.116-bit est. · params · source | No comparable published result | ~83 GB | 16-bit est. · params |
| Voxtral-Small-24B-250716-bit est. · params · source | No comparable published result | ~61 GB | 16-bit est. · params |
| Kimi-Audio-7B16-bit est. · params · source | No comparable published result | ~26 GB | 16-bit est. · params |
| Kimi-Audio-7B-Instruct16-bit est. · params · source | No comparable published result | ~26 GB | 16-bit est. · params |
| AuKConditional path · CUDA · source | No comparable published result | ~25 GB | Conditional path |
| AuK-FlashConditional path · CUDA · source | No comparable published result | ~25 GB | Conditional path |
33 more audio models in the catalog have no usable memory figure or parameter count for this profile. Browse Audio models
Embeddings
Embeddings · MTEB (Multilingual, v2) · mean(task)
0–100 · higher is better
- Color = publisher
| Qwen3-Embedding 8B8-bit · published params · source | 70.6 | ~12 GB | 8-bit · published params |
|---|---|---|---|
| Qwen3-Embedding 4B8-bit · published params · source | 69.5 | ~7 GB | 8-bit · published params |
| Harrier OSS v1 (0.6B)8-bit · published params · source | 69.0 | ~3 GB | 8-bit · published params |
| Qwen3-Embedding 0.6B8-bit · published params · source | 64.3 | ~3 GB | 8-bit · published params |
| Multilingual E5 Large Instruct8-bit · published params · source | 63.2 | ~3 GB | 8-bit · published params |
| EmbeddingGemma 28-bit · published params · source | 61.4 | ~3 GB | 8-bit · published params |
| EmbeddingGemma (300M)8-bit · published params · source | 61.1 | ~3 GB | 8-bit · published params |
| BGE-M38-bit · published params · source | 59.6 | ~3 GB | 8-bit · published params |
Runs well within 128 GB of unified memory · planning estimates, not measured speed
Measurements
Published tokens per second, exactly as measured; never estimated.
| gpt-oss 20B MXFP4 MoE | 83.43 ± 0.59 | MXFP4 MoE | CUDA | llama.cpp bench |
|---|---|---|---|---|
| gemma-3-4b-it-qat Q4_0 | 81.05 ± 0.20 | Q4_0 | llama.cpp CUDA | llama.cpp benches/dgx-spark · maintainer result |
| Qwen3.5 35B A3B MXFP4 | 63.13 | MXFP4 (MoE; 35B total / 3B active) | llama.cpp CUDA (SM121a, source build) | nabe2030/dense-27b-31b-dgx-spark (GitHub) · community benchmark |
| Qwen3 Coder 30B A3B Q8_0 | 61.06 ± 0.23 | Q8_0 | CUDA | llama.cpp bench |
| Llama 2 7B Q4_0 | 60.39 ± 0.05 | Q4_0 | Vulkan | llama.cpp Vulkan scoreboard · community result |
| gpt-oss-120b MXFP4 MoE | 58.72 ± 0.20 | MXFP4 MoE | llama.cpp CUDA | llama.cpp benches/dgx-spark · maintainer result |
| Nemotron-3-Nano-4B Q8_0 | 52.85 ± 0.12 | Q8_0 | llama.cpp CUDA | llama.cpp benches/nemotron · maintainer result |
| GLM-4.7-Flash Q8_0 | 48.68 ± 0.12 | Q8_0 | llama.cpp CUDA | llama.cpp benches/dgx-spark · maintainer result |
| Qwen3.8 27B NVFP4 + DSpark | 36.6 | NVFP4 + DSpark | SGLang | NVIDIA Developer Forum |
| Qwen2.5-Coder-7B-Q8_0 | 29.43 ± 0.02 | Q8_0 | llama.cpp CUDA | llama.cpp benches/dgx-spark · maintainer result |
| Nemotron 3 Super 120B A12B Q4_K | 19.94 ± 0.16 | Q4_K (reported as 'Q4_K' by llama-bench; GGUF from ggml-org) | llama.cpp CUDA | llama.cpp benches/nemotron · DGX Spark |
| Qwen3.8 27B NVFP4 + MTP | 18.5 | NVFP4 + MTP (3 speculative tokens) | vLLM official | NVIDIA Developer Forum |
| Qwen3.6 27B Q4_K_M | 11.85 ± 0.01 | Q4_K_M (15.65 GiB) | llama.cpp CUDA (SM121a, source build) | nabe2030/dense-27b-31b-dgx-spark (GitHub) · community benchmark |
| Qwen3.6 27B BF16 | 4.54 ± 0.00 | BF16 (50.10 GiB) | llama.cpp CUDA (SM121a, source build) | nabe2030/dense-27b-31b-dgx-spark (GitHub) · community benchmark |
Specifications
21 sourced fields; the source is named once per group, and a field from another source links to it. Official specifications · Memory specification checked: 2026-09-30 · manufacturer source · Sourced facts JSON
Markers: the best LiveBench-scored text model that runs well here, and the lightest text model that does not run well here but runs well with 256 GB. Q4 planning estimates, not measured fits.
Selected configuration
Latest check · Manufacturer source
- Memory bandwidth
- 273 GB/s
- CPU
- 20-core Arm (10 Cortex-X925 + 10 Cortex-A725)
- CPU cores
- 20
- GPU
- NVIDIA Blackwell
- Memory interface
- 256 bit
- CUDA cores
- 6,144
- Included external PSU rating
- 240 W
- AI performance (TOPS, manufacturer)
- 1,000 TOPS · FP4, sparse
Family specification
Latest check · Manufacturer source
- Storage options (family)
- 1 TB or 4 TB NVMe M.2
- Ports
- 4 × USB-C · HDMI 2.1a · 10Gb Ethernet · ConnectX-7
- Weight
- 1.2 kg
- Wi-Fi
- Wi-Fi 7
- Bluetooth
- Bluetooth 5.4
- Dimensions (cm)
- 15.0 × 15.0 × 5.05 (L × W × H)
128 GB coherent unified memory. Check ARM64 runtime and checkpoint support. Capacity alone does not establish throughput or compatibility.
Photo: Official manufacturer image. · Manufacturer source
Memory bandwidth is not measured model speed; verify runtime and checkpoint compatibility. Hardware specification reviewed: 2026-10-18.
How estimates work
- Candidates
- Reviewed pages use published parameter counts; automatic text candidates use official repository tensor counts (packed quantized weights and unsupported architectures excluded).
- Planning
- Q4 weights + 20% allowance + reserve: 25% of unified or system memory (at least 2 GB, at most 8 GB; 8 GB here) or 2 GB per dedicated GPU. Adjustable in “Adjust the estimate”. No KV cache is calculated from context.
- Fit tiers
- Runs well: the estimate fits the fast memory (VRAM, or unified memory within the macOS GPU share). Runs slowly: within the machine's memory but outside the fast part; a graphics card offloads layers to PC RAM (the listed RAM, or 32 GB assumed). Doesn't run: above both. Tiers are planning estimates, never a speed.
- Vision
- Vision-language models use the Q4 planning estimate from the parameter count (published by the model page, or read from Hugging Face). Classic vision models (classify, detect, segment) use a 16-bit estimate. No vision benchmark is tracked, so rows carry no score.
- Media and embeddings
- Media rows use documented execution paths (exact hardware or platform) and publisher minimums; a published requirement above this memory moves a model to the next tier. Without either, a 16-bit estimate from the published parameter count (× 1.2 + 2 GB), labelled as such. Embeddings use 8-bit weights + 20% + 2 GB.
Local AI basics: parameters, Q4 and memory · Full model catalog · Data and methodology
Alternatives
Side-by-side comparisons, then other machines on the same memory ladder (unified memory); each page lists the models that fit it.
Compare with… 3
Same chip · NVIDIA GB10 5
Also 128 GB 6
- Framework Desktop · Ryzen AI Max+ 395 · 128 GB
- HP Z2 Mini G1a · Ryzen AI Max+ PRO 395 · 128 GB unified
- Mac Studio M5 Max · 128 GB unified
- MacBook Pro 14-inch · M5 Max · 18-core CPU / 40-core GPU · 128 GB unified
- MacBook Pro 16-inch · M5 Max · 18-core CPU / 40-core GPU · 128 GB unified
- ROG Flow Z13 (2025) · Ryzen AI Max+ 395 · 128 GB unified
