What LLMs can a DGX Spark (128 GB) run?
A DGX Spark (128 GB), with 120 GB usable by the GPU, runs 49 of the 69 open models tracked here at Q4_K_M with 8K context; the largest is Qwen3.8 Flash Next (180.0B parameters, 112 GB). It reads memory at 273 GB/s. 2 more fit at a lower precision. Each fit leaves at least 0.5 GB free on the DGX Spark (128 GB) (why a card does not give a model all its memory).
Models that fit on one DGX Spark (128 GB)
The most precise weights that still fit with 8K tokens of context and one request, the memory that takes, the longest context at that precision and the writing speed. Bigger models come first. gpt-oss-120b (67.7 GB, 25–42 tokens/s on the DGX Spark (128 GB)) and gpt-oss-20b (14.8 GB, 30–51 tokens/s on the DGX Spark (128 GB)) are listed at MXFP4 only, as published.
| Model | Best precision | Memory | Longest context | Tokens/s |
|---|---|---|---|---|
| MiniMax M2.7 228.7B, 10B active | GGUF Q3_K_M | 117 GB | 16K | 12–19 |
| Step 3.7 Flash 196B-A11B 201.4B, 11B active | GGUF Q3_K_M | 102 GB | 256K (full) | 14–23 |
| Qwen3.8 Flash Next 180.0B, 6B active | GGUF Q4_K_M | 112 GB | 256K (full) | 21–35 |
| Mistral Medium 3.5 128B 127.7B | GGUF Q6_K | 111 GB | 30K | 1.4–1.9 |
| Qwen3.5 122B-A10B 125.1B, 10B active | GGUF Q6_K | 106 GB | 256K (full) | 9.6–16 |
| Nemotron 3.5 Super VL 120B-A12B 124.5B, 12B active | GGUF Q6_K | 105 GB | 1M (full) | 8.2–14 |
| Nemotron 3 Super 120B-A12B 123.6B, 12B active | GGUF Q6_K | 104 GB | 256K (full) | 8.2–14 |
| gpt-oss-120b 116.8B, 5.1B active | As published (MXFP4) | 67.7 GB | 128K (full) | 25–42 |
| AliceAI Foundation 80B-A3B 81.3B, 3B active | GGUF Q8_0 | 89.2 GB | 256K (full) | 23–39 |
| Qwen3-Coder-Next 79.7B, 3B active | GGUF Q8_0 | 87.4 GB | 256K (full) | 23–39 |
| Kolibri-1 78.1B, 3.5B active | GGUF Q8_0 | 85.8 GB | 256K (full) | 20–34 |
| Llama 3.1 70B 70.6B | GGUF Q8_0 | 80.0 GB | 122K | 1.9–2.6 |
| K2-Horizon MoVA 36B-A4B 37.4B, 4B active | FP16 / BF16 | 78.9 GB | 204K | 8.4–14 |
| Ornith 1.5 35B-A3B 36.0B, 3B active | FP16 / BF16 | 74.3 GB | 256K (full) | 13–22 |
| Qwen3.6 35B-A3B 36.0B, 3B active | FP16 / BF16 | 74.3 GB | 256K (full) | 13–22 |
| Ornith 1.0 35B 35.1B, 3B active | FP16 / BF16 | 72.6 GB | 256K (full) | 13–22 |
| LLM-jp-4.1 32B-A3B Thinking 32.1B, 3.8B active | FP16 / BF16 | 66.9 GB | 64K (full) | 9.9–17 |
| Nemotron 3 Nano 30B-A3B 31.6B, 3.5B active | FP16 / BF16 | 65.3 GB | 256K (full) | 11–19 |
| Gemma 4 31B 31.3B | FP16 / BF16 | 66.6 GB | 256K (full) | 2.3–3.2 |
| GLM-4.7 Flash 31.2B, 3B active | FP16 / BF16 | 64.9 GB | 198K (full) | 12–21 |
| Xing 4.0 29B-A4B 31.2B, 4B active | FP16 / BF16 | 64.8 GB | 256K (full) | 9.6–16 |
| Qwen3-Coder 30B-A3B 30.5B, 3.3B active | FP16 / BF16 | 63.9 GB | 256K (full) | 11–18 |
| Qwen3 30B-A3B 30.5B, 3.3B active | FP16 / BF16 | 63.9 GB | 40K (full) | 11–18 |
| Muse Glimmer 30B 29.8B | FP16 / BF16 | 61.7 GB | 128K (full) | 2.5–3.4 |
| Granite 4.2 30B 29.3B | FP16 / BF16 | 62.7 GB | 128K (full) | 2.5–3.4 |
| ThinkingCap Qwen3.8 27B 27.8B | FP16 / BF16 | 58.0 GB | 256K (full) | 2.7–3.6 |
| Qwen3.6 27B 27.8B | FP16 / BF16 | 58.0 GB | 256K (full) | 2.7–3.6 |
| Qwen3.8 27B 27.8B | FP16 / BF16 | 58.0 GB | 256K (full) | 2.7–3.6 |
| Hemmingway-1 27B 27.3B | FP16 / BF16 | 57.0 GB | 256K (full) | 2.7–3.7 |
| Gemma 4 26B-A4B 25.8B, 4B active | FP16 / BF16 | 53.9 GB | 256K (full) | 9.5–16 |
| gpt-oss-20b 20.9B, 3.6B active | As published (MXFP4) | 14.8 GB | 128K (full) | 30–51 |
| Gemma 4 12B 12.0B | FP16 / BF16 | 25.7 GB | 256K (full) | 6.1–8.3 |
| ZDTaichu 5.0 9B 9.8B | FP16 / BF16 | 20.8 GB | 128K (full) | 7.5–10 |
| Ornith 1.5 9B 9.7B | FP16 / BF16 | 20.6 GB | 256K (full) | 7.6–10 |
| Qwen3.5 9B 9.7B | FP16 / BF16 | 20.6 GB | 256K (full) | 7.6–10 |
| LensVLM 9B 9.4B | FP16 / BF16 | 20.1 GB | 256K (full) | 7.8–11 |
| Ornith 1.0 9B 9.4B | FP16 / BF16 | 20.1 GB | 256K (full) | 7.8–11 |
| MiMo V2.6 Distill Qwen 9B 9.4B | FP16 / BF16 | 20.1 GB | 256K (full) | 7.8–11 |
| Bespoke Nimble 9B 9.0B | FP16 / BF16 | 19.1 GB | 8K (full) | 8.2–11 |
| Granite 4.2 8B 8.8B | FP16 / BF16 | 19.9 GB | 128K (full) | 7.8–11 |
| LFM2.5 8B-A1B 8.5B, 1.5B active | FP16 / BF16 | 18.0 GB | 128,000 (full) | 25–43 |
| Qwen3 8B 8.2B | FP16 / BF16 | 18.5 GB | 40K (full) | 8.4–12 |
| Llama 3.1 8B 8.0B | FP16 / BF16 | 18.1 GB | 128K (full) | 8.6–12 |
| Gemma 4 E4B 8.0B | FP16 / BF16 | 17.1 GB | 128K (full) | 9.2–13 |
| Ling 3.0 Tiny 7.9B-A1.3B 7.9B, 1.3B active | FP16 / BF16 | 16.7 GB | 128K (full) | 29–50 |
| LightOnOCR-3 4B 4.5B | FP16 / BF16 | 10.1 GB | 256K (full) | 16–22 |
| Spark-X2.5 4B 4.1B | FP16 / BF16 | 9.35 GB | 1M (full) | 17–23 |
| Nemotron 3 Nano 4B 4.0B | FP16 / BF16 | 8.78 GB | 256K (full) | 18–25 |
| Granite 4.2 3B 3.7B | FP16 / BF16 | 8.69 GB | 128K (full) | 18–25 |
| MiniCPM5 2B 2.5B | FP16 / BF16 | 6.02 GB | 128K (full) | 27–37 |
| Limite 1B Violetto 1.0B | FP16 / BF16 | 2.79 GB | 128K (full) | 61–88 |
Too big for one DGX Spark (128 GB) at 8K context
How many DGX Spark (128 GB) cards these need at Q4_K_M with 8K context and 0.5 GB to spare, in one tensor-parallel group; some need more than 8.
- DeepSeek V4 Flash: 2 cards
- DeepSeek V4 Flash 0731: 2 cards
- MiMo V2.6 Flash: 2 cards
- IQuest-Q1 320B-A15B: 2 cards
- GLM-5.3 Flash: 2 cards
- MiniMax M3: 4 cards
- Ling 3.1 Flash 560B-A25B: 4 cards
- DeepSeek V3 / R1: 4 cards
- DeepSeek V3.2: 4 cards
- GLM-5.3: 4 cards
- GLM-5.2: 4 cards
- DeepSeek V4.1 Flash 1017: 4 cards
- DeepSeek V4.1 Flash: 4 cards
- Hy4 Preview 770B-A49B: 8 cards
- MiMo V2.6 Pro: 8 cards
- DeepSeek V4 Pro: more than 8 cards
- Qwen3.8 2.4T-A95B: more than 8 cards
- Kimi K3: more than 8 cards
Best models for DGX Spark (128 GB) with room for long context
The biggest models that still fit at Q4_K_M with 32,768 tokens of context (or their whole window, if shorter), the memory left over, the longest context the card takes at that precision and the writing speed with that context in the cache, from 273 GB/s of bandwidth.
| Model | Context | Memory | Left over | Longest context | Tokens/s |
|---|---|---|---|---|---|
| Qwen3.8 Flash Next 180.0B | 32K | 113 GB | 7.11 GB | 256K (full) | 18–30 |
| Mistral Medium 3.5 128B 127.7B | 32K | 91.8 GB | 28.2 GB | 105K | 1.7–2.3 |
| Qwen3.5 122B-A10B 125.1B | 32K | 78.9 GB | 41.1 GB | 256K (full) | 12–20 |
| Nemotron 3.5 Super VL 120B-A12B 124.5B | 32K | 77.9 GB | 42.1 GB | 1M (full) | 11–18 |
| Nemotron 3 Super 120B-A12B 123.6B | 32K | 77.4 GB | 42.6 GB | 256K (full) | 11–18 |
llama-server commands for one DGX Spark (128 GB)
GGUF repos checked 2026-09-29; -c is the longest context with 0.5 GB free on the DGX Spark (128 GB). No -ngl: llama.cpp’s -fit, on by default since b7440, places the layers.
Mistral Medium 3.5 128B, Q4_K_M with 105K tokens: 1.3–1.8 tokens/s on the DGX Spark (128 GB)
llama-server -hf unsloth/Mistral-Medium-3.5-128B-GGUF:Q4_K_M -c 107520Q4_K_M/
Qwen3.5 122B-A10B, Q4_K_M with 256K tokens: 6.5–11 tokens/s on the DGX Spark (128 GB)
llama-server -hf unsloth/Qwen3.5-122B-A10B-GGUF:Q4_K_M -c 262144Q4_K_M/
How fast the DGX Spark (128 GB) writes as the context fills
Tokens per second at Q4_K_M (gpt-oss-120b at MXFP4) for the biggest models that fit with 32K tokens, with 8K and 32K tokens in the cache and at the longest context the card holds, and at Q8_0 with 8K where that fits. Each token reads the active weights and the whole cache once, so the speed follows the 273 GB/s of bandwidth and falls as the cache grows. The ceiling is that bandwidth divided by the bytes read per token, which no runtime reaches; the reply time is for 1,000 new tokens with 32K (or the longest context) in the cache, prompt processing aside.
| Model | 8K | 32K | Longest | At the longest | Q8_0 at 8K | Bandwidth ceiling at 8K | 1,000-token reply at 32K |
|---|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next | 21–35 | 18–30 | 256K | 8.0–13 | Does not fit | 71 | 33–56 s |
| Mistral Medium 3.5 128B | 1.9–2.5 | 1.7–2.3 | 105K | 1.3–1.8 | Does not fit | 3 | 436–595 s |
| Qwen3.5 122B-A10B | 13–22 | 12–20 | 256K | 6.5–11 | Does not fit | 44 | 51–85 s |
| Nemotron 3.5 Super VL 120B-A12B | 11–18 | 11–18 | 1M | 5.1–8.6 | Does not fit | 37 | 56–93 s |
| Nemotron 3 Super 120B-A12B | 11–18 | 11–18 | 256K | 8.6–14 | Does not fit | 37 | 56–93 s |
| gpt-oss-120b | 25–42 | 19–33 | 128K | 10–18 | MXFP4 only | 86 | 30–51 s |
| AliceAI Foundation 80B-A3B | 38–65 | 30–51 | 256K | 9.8–16 | 23–39 | 135 | 20–33 s |
| Qwen3-Coder-Next | 38–65 | 30–51 | 256K | 9.8–16 | 23–39 | 135 | 20–33 s |
| Kolibri-1 | 33–56 | 27–46 | 256K | 11–18 | 20–34 | 115 | 22–37 s |
| Llama 3.1 70B | 3.3–4.5 | 2.8–3.8 | 128K | 1.7–2.4 | 1.9–2.6 | 6 | 261–357 s |
Near misses on DGX Spark (128 GB)
Models that miss at Q4_K_M with 32,768 tokens of context (or their whole window), or leave under 0.5 GB free there, but fit with a shorter context or 3-bit weights, or are short by at most a quarter of the memory and fit with MoE offload. Smallest shortfall first.
| Model | Short by at Q4_K_M | Fits instead |
|---|---|---|
| Step 3.7 Flash 196B-A11B 201.4B | 7.10 GB | GGUF Q3_K_M with 32K, 103 GB |
| MiniMax M2.7 228.7B | 30.8 GB | GGUF IQ3_XXS with 32K, 106 GB |
Run locally or rent a DGX Spark (128 GB)?
getdeploying.com lists no on-demand rental of DGX Spark (128 GB) (checked ). Among the GPUs this site tracks, the cheapest to rent with at least 120 GB is the H200 at a median $4.46 an hour, which holds 50 of the 69 models at Q4_K_M with 8K context against 49 here; every $1,000 equals about 224 hours of it.
How these numbers are worked out
Memory is the weights at each precision plus the KV cache for 8,192 tokens in FP16 and runtime overhead (0.5 GB plus 10%), computed from each model's files on Hugging Face with the LLM VRAM Calculator. Speed is estimated from memory bandwidth and the parameters read per token with the LLM Speed Calculator. Both tools take any other model from Hugging Face.
Nearby GPUs and setups
How many of the tracked models each one holds at Q4_K_M with 8K context, against 49 here.
| Setup | How it relates | Memory | Bandwidth | Models at Q4_K_M |
|---|---|---|---|---|
| 4× RTX 3090 | Next size down | 96 GB | 3,744 GB/s | 48 (−1) |
| H200 | Next size up | 141 GB | 4,800 GB/s | 50 (+1) |
Model numbers read from Hugging Face on .