What can a DGX Spark run? Memory limits and recipes
One DGX Spark has 128 GB LPDDR5x coherent unified system memory. This catalog reads that as 128 GiB and reserves 16 GiB for the OS, so the model budget is 112 GiB. The largest model in our list that runs cleanly on one unit at 8,192 tokens is gpt-oss-120b with MXFP4 (61.7 GiB–63.6 GiB). The memory is shared with the CPU and is not dedicated VRAM, and several Sparks do not form one pool.
One unit: model by model
| Model | Best file that runs | Total at 8,192 | Result | Smallest file |
|---|---|---|---|---|
| gpt-oss-120b | MXFP4 | 61.7 GiB–63.6 GiB | Comfortable | MXFP4, 60.8 GiB |
| Qwen2.5-72B-Instruct | Q8_0 | 74.8 GiB–76.8 GiB | Comfortable | Q4_K_M, 41.0 GiB |
| Qwen3-235B-A22B | none | n/a | No | Q4_K_M, 132.4 GiB |
| Llama 3.3 70B Instruct | none | n/a | No | BF16, 131.4 GiB |
| Gemma 3 27B IT | BF16 | 52.8 GiB–54.7 GiB | Comfortable | Q4_0, 16.0 GiB |
| Qwen3-32B | BF16 | 63.6 GiB–65.6 GiB | Comfortable | Q4_K_M, 18.4 GiB |
| Qwen3-30B-A3B | BF16 | 58.2 GiB–60.1 GiB | Comfortable | Q4_K_M, 17.3 GiB |
| Mistral-Small-24B-Instruct-2501 | BF16 | 45.8 GiB–47.7 GiB | Comfortable | Q4_K_M, 13.3 GiB |
| gpt-oss-20b | MXFP4 | 13.6 GiB–15.5 GiB | Comfortable | MXFP4, 12.8 GiB |
| Llama 3.1 8B Instruct | BF16 | 16.6 GiB–18.5 GiB | Comfortable | BF16, 15.0 GiB |
Net budget 112 GiB, llama.cpp-style estimate, one sequence, fp16 KV cache, typical extras (calculator 0.3.0, dataset 2026-10-02.1). "No" means even the smallest file is over budget. These are memory estimates; no model here was timed on a Spark by this site, and no tokens-per-second figure is given.
Deployment recipes on DGX Spark
Recipes are third-party setups with their sources and check dates recorded. Evidence labels come from each recipe record. Open a recipe for the artifact bytes, runtime version and what was and was not tested.
- DeepSeek-V4.1-Flash on 2× DGX Spark (EXL3 2.9 bpw) — 2 Sparks, vLLM, maintainer-measured, checked 2026-09-23. 2-node DGX Spark setup serving DeepSeek-V4.1-Flash with vLLM tensor-parallel size 2 over ConnectX-7 RoCE, EXL3 2.9 bpw mul1 weights, and native NVMe-resident Engram tables.
- DeepSeek-V4.1-Flash on 3× DGX Spark (SGLang MXFP4 / FP8) — 3 Sparks, SGLang, maintainer-measured, checked 2026-09-23. 3-node DGX Spark cluster running native MXFP4 experts and FP8 dense weights with SGLang TP=3 over ConnectX-7 RoCE triangle, with NVMe-resident Engram tables.
- DeepSeek-V4.1-Flash on 4× DGX Spark (1M Long-Context Setup) — 4 Sparks, SGLang, maintainer-measured, checked 2026-09-23. 4-node DGX Spark cluster running native MXFP4 MoE and FP8 dense weights with SGLang TP=4, verified 1M context needle test, and 8M pinned KV token pool.
- DeepSeek-V4-Flash-0731 on 2× DGX Spark (DSpark Speculative) — 2 Sparks, vLLM, maintainer-measured, checked 2026-09-23. 2-node DGX Spark deployment for DeepSeek-V4-Flash-Vision-Exp running vLLM TP=2 with DSpark speculative decoding and nvfp4_ds_mla KV cache.
- GLM-5.3-Flash on 2× DGX Spark (EXL3 4 bpw) — recipe page with the same fields.
Community profiles
Similar 128 GB unified machines
| Device | Capacity | Default reserve | Model budget |
|---|---|---|---|
| NVIDIA DGX Spark (GB10, 128 GB unified) | 128 GiB | 16 GiB | 112 GiB |
| Ryzen AI Max+ 395 | 128 GiB | 32 GiB | 96 GiB |
| Apple M4 Max 128GB unified (MacBook Pro) | 128 GiB | 12 GiB | 116 GiB |
Reserves differ because the sources differ. For the Ryzen AI Max+ 395, AMD lists up to 96 GB of Variable Graphics Memory on the 128 GB configuration, so 32 GiB stays with the system. The Apple record assumes a raised GPU limit; at the default macOS share a 128 GB Mac gives the GPU about 96 GB, see the 128 GB Mac page. Memory bandwidth differs by machine, and this site does not convert it into speed.
Data sources
- NVIDIA DGX Spark product page — 128 GB LPDDR5x coherent unified system memory. This is not dedicated VRAM. Default reserve deducts OS and other processes. Multiple Sparks are not one pool.
- NVIDIA DGX Spark hardware docs — 128 GB LPDDR5x unified memory on the GB10 Superchip.
- AMD Ryzen AI Max+ 395 specifications — System Memory Type 256-bit LPDDR5x; Max. Memory 128 GB; Max Memory Speed LPDDR5x-8000; Graphics Model Radeon 8060S; Graphics Core Count 40; Default TDP 55 W, cTDP 45-120 W
- AMD FAQ: Variable Graphics Memory — The 128 GB Ryzen AI Max+ 395 machine can dedicate up to 96 GB to the iGPU with Variable Graphics Memory (BIOS-level; Adrenalin 25.8.1 or later on Windows). That memory is removed from system RAM.
- Recipe records — Each recipe lists its own source repository, commit and check date