LLMBOTTLENECK.COM
NVIDIA / unified

NVIDIA DGX Spark

128 GB decides what fits. 273 GB/s decides how fast it runs once it does.

Open in the calculator →Best models for 128 GB, on every card that size →

Find one: Amazon ↗ · eBay (new and used) ↗Store links may pay us a commission. They never decide which card is suggested — the memory arithmetic does.

Quick answer
The NVIDIA DGX Spark fits 248 of 320 sized models entirely in its 128 GB; the largest widely used one is Qwen3.8-Flash-Next (111.6 GB at Q4_K_M). Memory decides what fits; its 273 GB/s of bandwidth decides how fast it answers.

At 8,192 tokens of context with the whole model in device memory. Speeds are estimates from memory bandwidth, not benchmarks run on this card.

NVIDIA DGX Spark · 128 GB · 273 GB/s · Q4_K_M where published · 8,192 tokensSpecificationpublished dataDecode speedestimate

248 of 320 fit entirely

248 of the 320 models the engine can size fit entirely in device memory at Q4_K_M where it is published, otherwise the nearest published format, and 8,192 tokens. Unified memory is one pool, so there is no second memory tier to spill into; the remaining 72 do not run at this context.

78%of the models the engine can size fit entirely
128GB of device memory
273GB/s memory bandwidth
0run with system memory
72do not run at 8,192 tokens

Featured models · Q4_K_M at 8,192 tokens, including offload

ModelNeedsVerdictDecodeCalculator
Qwen3.5-2B2.3B parameters2.32 GBfits~150 tok/sOpen →
NVIDIA-Nemotron-3-Nano-4B-BF164.0B parameters3.46 GBfits~70 tok/sOpen →
Qwen3-VL-8B-Instruct8.8B parameters7.04 GBfits~32 tok/sOpen →
Qwen3.5-9B9.7B parameters7.05 GBfits~38 tok/sOpen →
gemma-4-12B-it12B parameters9.54 GBfits~23 tok/sAbout reading paceOpen →
gpt-oss-20b21B parameters13.8 GBfits~74 tok/sOpen →
Qwen3.8-27B28B parameters18.6 GBfits~13 tok/sAbout reading paceOpen →
Qwen3.6-27B28B parameters18.6 GBfits~13 tok/sAbout reading paceOpen →
Kimi-Linear-48B-A3B-Instruct49B parameters30.4 GBfits~99 tok/sOpen →
Qwen3.8-Flash-Next180B parameters111.6 GBfits~52 tok/sOpen →
DeepSeek-V4-Flash-Vision-Exp305B parameters188.1 GBshort by 60.1 GBdoes not runOpen →
GLM-5.3-Flash321B parameters198.3 GBshort by 70.3 GBdoes not runOpen →

These examples are selected from prominent labs using the catalogue’s latest Hugging Face 30-day downloads and repository-creation freshness signal, with newer releases guaranteed a place. Offloaded rows assume 32 GB of system RAM, and a speed is only shown for a row that runs. Fitting in memory is not the same as loading: whether the runtime and version you have supports each architecture and format on this machine has not been tested here. Each “Open” link carries the same model, format, context and RAM into the calculator.

Fits entirely in NVIDIA DGX Spark memory — 248 of 320 sized models

Every model that fitswhole model resident · offload off

The most demanding model that fits is Step-3.5-Flash at Q4_K_M: 127.0 GB of the 128 GB, leaving 1.03 GB spare.

Fully resident at 8,192 tokens (or the model’s own maximum, where that is shorter), offload off, at 273 GB/s. Each row is a run of the engine for this configuration; the rows start with current, prominent releases and “fits” is memory, not a tested runtime. Older or less prominent models remain available through this search and “Show all”.

Showing 40 of 248 models that fit.

ModelFormatNeedsSpareDecodeDownloads / 30dCalculator
Qwen3.8-27BQwen · 28B paramsQ4_K_M18.6 GB109.4 GB~13 tok/sAbout reading pace6.9MOpen →
Qwen3.8-Flash-NextNEWQwen · 180B paramsQ4_K_M111.6 GB16.4 GB~52 tok/s1.4MOpen →
gemma-4-26B-A4B-itGoogle · 26B paramsQ4_K_M16.8 GB111.2 GB~70 tok/s13MOpen →
gemma-4-31B-itGoogle · 31B paramsQ4_K_M22.2 GB105.8 GB~9.0 tok/sSlow9.9MOpen →
Qwen3.5-9BQwen · 9.7B paramsQ4_K_M7.05 GB121.0 GB~38 tok/s9MOpen →
Qwen3.5-4BQwen · 4.7B paramsQ4_K_M3.98 GB124.0 GB~69 tok/s7.8MOpen →
Qwen3.6-35B-A3BQwen · 36B paramsQ4_K_M23.1 GB104.9 GB~100 tok/s3.3MOpen →
gemma-4-12B-itGoogle · 12B paramsQ4_K_M9.54 GB118.5 GB~23 tok/sAbout reading pace1.9MOpen →
Qwen3.6-27BQwen · 28B paramsQ4_K_M18.6 GB109.4 GB~13 tok/sAbout reading pace2.5MOpen →
Qwen3.5-2BQwen · 2.3B paramsQ4_K_M2.32 GB125.7 GB~150 tok/s4.9MOpen →
gemma-4-E4B-itGoogle · 8.0B paramsQ4_K_M5.86 GB122.1 GB~61 tok/s4.4MOpen →
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16NVIDIA · 32B paramsQ4_K_M20.3 GB107.7 GB~9.5 tok/sSlow530KOpen →
NVIDIA-Nemotron-3-Nano-4B-BF16NVIDIA · 4.0B paramsQ4_K_M3.46 GB124.5 GB~70 tok/s3.4MOpen →
gemma-4-E2B-itGoogle · 5.1B paramsQ4_K_M4.02 GB124.0 GB~162 tok/s3MOpen →
Qwen3.5-0.8BQwen · 873M paramsQ4_K_M1.46 GB126.5 GB~330 tok/s2.6MOpen →
Qwen3-VL-8B-InstructQwen · 8.8B paramsQ4_K_M7.04 GB121.0 GB~32 tok/s14.6MOpen →
Qwen3.5-27BQwen · 28B paramsQ4_K_M18.6 GB109.4 GB~13 tok/sAbout reading pace1.9MOpen →
North-Micro-Vision-InstructCohereLabs · 2.5B paramsQ4_K_M2.99 GB125.0 GB~99 tok/s180.7KOpen →
Qwen3.5-35B-A3BQwen · 36B paramsQ4_K_M23.1 GB104.9 GB~100 tok/s1.6MOpen →
NVIDIA-Nemotron-3-Super-120B-A12B-BF16NVIDIA · 124B paramsQ4_K_M76.9 GB51.1 GB~2.4 tok/sVery slow1.2MOpen →
granite-4.2-8bIBM · 8.8B paramsQ4_K_M7.49 GB120.5 GB~29 tok/sAbout reading pace128.4KOpen →
GLM-4.7-Flashzai-org · 31B paramsQ4_K_M20.4 GB107.6 GB~9.5 tok/sSlow1.8MOpen →
granite-4.1-3bIBM · 3.4B paramsQ4_K_M3.72 GB124.3 GB~67 tok/s521.8KOpen →
LFM2.5-2.6BLiquidAI · 2.7B paramsQ4_K_M2.61 GB125.4 GB~107 tok/s108.1KOpen →
Qwen3-VL-4B-InstructQwen · 4.4B paramsQ4_K_M4.72 GB123.3 GB~50 tok/s3.5MOpen →
granite-4.1-30bIBM · 29B paramsQ4_K_M20.4 GB107.6 GB~9.4 tok/sSlow301.5KOpen →
Qwen3.5-122B-A10BQwen · 125B paramsQ4_K_M77.9 GB50.1 GB~37 tok/s512.6KOpen →
Qwen3-VL-2B-InstructQwen · 2.1B paramsQ4_K_M3.05 GB125.0 GB~91 tok/s2.8MOpen →
granite-4.2-3bIBM · 3.7B paramsQ4_K_M3.72 GB124.3 GB~67 tok/s50.6KOpen →
LFM2.5-230MLiquidAI · 230M paramsQ4_K_M1.05 GB126.9 GB~749 tok/s88.6KOpen →
Qwen3-0.6BQwen · 752M paramsQ4_K_M2.22 GB125.8 GB~139 tok/s29.7MOpen →
Qwen3-Coder-NextQwen · 80B paramsQ4_K_M50.0 GB78.0 GB~67 tok/s596.3KOpen →
gpt-oss-20bOpenAI · 21B paramsQ4_K_M13.8 GB114.2 GB~74 tok/s6.6MOpen →
granite-4.2-30bIBM · 29B paramsQ4_K_M20.7 GB107.3 GB~9.4 tok/sSlow35.8KOpen →
granite-4.1-8bIBM · 8.8B paramsQ4_K_M7.49 GB120.5 GB~29 tok/sAbout reading pace179.3KOpen →
NVIDIA-Nemotron-3-Nano-30B-A3B-BF16NVIDIA · 32B paramsQ4_K_M20.3 GB107.7 GB~9.5 tok/sSlow875.2KOpen →
gpt-oss-120bOpenAI · 117B paramsQ4_K_M71.9 GB56.1 GB~52 tok/s4.5MOpen →
LFM2.5-VL-3BLiquidAI · 3.1B paramsQ4_K_M2.85 GB125.1 GB~107 tok/s26.7KOpen →
MiMo-V2.6-Distill-Qwen-9BNEWXiaomiMiMo · 9.4B paramsQ4_K_M6.90 GB121.1 GB~38 tok/s14.6KOpen →
Qwen3-4B-Instruct-2507Qwen · 4.0B paramsQ4_K_M4.72 GB123.3 GB~50 tok/s3.7MOpen →

Sized at the format most people actually download, not at FP16. The catalogue ordered by downloads →

Find the NVIDIA DGX Spark: Amazon ↗ · eBay (new and used) ↗Store links may pay us a commission. They never decide which card is suggested — the memory arithmetic does.

The DGX Spark, model by model

One page per model: whether it fits this device, at which formats, and how fast.

Under the hood

The specification behind every figure

What the manufacturer publishes for this device, and the pages it was read from.

Manufacturer specification

Memory scopeunified-system
Capacity128 GB
Published options128 GB
Bandwidth273 GB/s
Memory typeLPDDR5x
Bus widthNot published
FP32 peakNot published
Dense matrix peakNot published without sparsity
power supply240 W

Source ledger

NVIDIA DGX Spark ↗
Retrieved 2026-08-31

Caveats

  • The published AI performance figure is 1 PFLOP at FP4 on the Blackwell tensor cores. It is not a general-purpose throughput figure and is not used to price prefill here.
  • Memory is coherent unified system memory shared with a 20-core Arm CPU; the operating system's own use is not subtracted.
  • No dense matrix throughput is catalogued for this device: NVIDIA publishes a sparse FP4 petaFLOP figure for this system and no dense matrix throughput. Time to first token is withheld rather than priced against a sparsity figure.

Machines built on the NVIDIA DGX Spark

ASUS Ascent GX10

Sold with 128 GB

  • platform and memory: “CPU ARM v9.2-A CPU (GB10) … Integrated Graphics NVIDIA Blackwell GPU (GB10, integrated) … Memory 128 GB LPDDR5x, unified system memory” — www.asus.com, read 2026-09-13
  • ASUS does not state a memory bandwidth on this page. The engine uses the figure NVIDIA publishes for its own GB10 system (DGX Spark); that the ASUS board reaches the same figure is an inference from the shared GB10 platform, not an ASUS statement.
  • Cooling, firmware and networking differ from NVIDIA's DGX Spark and are not modelled.

Size a model on the 128 GB ASUS Ascent GX10 →

Dell Pro Max with GB10

Sold with 128 GB

  • platform and memory: “Dell Pro Max with GB10 Model: FCM1253 … NVIDIA GB10 Grace CPU (10 Cortex-X925 + 10 Cortex-A725 cores) NVIDIA DGX OS 7 NVIDIA GB10 Blackwell GPU 128GB LPDDR5X” — www.dell.com, read 2026-09-13
  • Dell's US configurator page does not state a memory bandwidth. The engine uses the figure NVIDIA publishes for its own GB10 system (DGX Spark); that Dell's board reaches the same figure is an inference from the shared GB10 platform, not a Dell statement.
  • Price and availability shown by Dell are regional and dated, and are not recorded.

Size a model on the 128 GB Dell Pro Max with GB10 →

Published capacity is a hardware ceiling, not guaranteed free runtime memory. The calculator shows the runtime reserve separately rather than folding it into a single number.

Run the diagnostic on the NVIDIA DGX Spark →
llmbottleneck
catalogue 2026-10-03models 327devices 135