What LLMs fit in 128GB VRAM? (Oct 2026)
39 models on our list fit in 128 GB of graphics memory. The largest by parameter count is DeepSeek V4.1 Flash 1017 at UD-IQ1_S, Engram table on SSD, 123.18 GB (Unsloth guide: 1-bit needs 128–145 GB RAM+VRAM with the Engram table on SSD via mmap; not tested by us. Our estimate leaves out the conversation memory for this model: real use needs more than the figures shown). We keep 0.6 GB free on one graphics card, so 127.4 GB is usable. Each total adds conversation memory for 8,192 tokens (or the model's own maximum, when that is shorter) and 0.8 GB for the software. The total is an estimate.
Updated
A self-reported pair of 128 GB Ryzen AI Max+ 395 boxes (Strix Halo), not a localLLMrig measurement: @c4software runs Qwen3.8-27B on bigchuck with llama.cpp and on a Minisforum MS-S1 MAX (silentchuck) with gufo. Since gufo 0.11.0 his gufo measurements run on silentchuck, a Minisforum MS-S1 MAX with a Ryzen AI Max+ 395 and 128 GB, not on bigchuck. Read the interview. Minisforum MS-S1 Max 128GB.
See also Xijing (Thomas) Wang's MacBook Pro M5 Max 128 GB (about 107 GB GPU-usable), serving Qwen3.8-27B at 5-bit with oMLX. A 64K prompt prefill takes about 110 seconds. Self-reported. Read Xijing (Thomas) Wang's interview. Mac Studio M5 Max 128GB.
See also James Meadlock's GhettoRak: eight NVIDIA DGX Spark nodes at 128 GB unified memory each (1,024 GB across the cluster), running GLM-5.3 at NVFP4. Self-reported. Read James Meadlock's interview. NVIDIA DGX Spark 128GB.
See also Valérian de Thézan de Gaussan's Framework Desktop (AMD Ryzen AI Max+ 395, Strix Halo, 128 GB shared LPDDR5X, ~111.56 GiB exposed to Vulkan). Mistral-Small-4-119B-2603-UD-Q4_K_M decoded at 32.72 ± 0.020 tok/s on llama.cpp Vulkan. Self-reported. Read Valérian de Thézan de Gaussan's interview. Framework Desktop 128GB.
Own a 128 GB GPU? Get a 1-line alert when a new top model fits 127.4 GB usable VRAM
39 models on our catalog fit 128 GB today. Get our weekly 1-line alert when a new open-weight LLM lands inside 127.4 GB at Q4_K_M or when 128 GB+ GPU prices drop >10% — zero spam.
Models that fit
Ranked by parameter count, largest first. The best quant is the highest-quality option that fits in 127.4 GB. A measured download is used when we have one. Otherwise the figure is Q4, Q5, Q8 or FP16 from our estimate. The gigabytes are the weights plus conversation memory for 8,192 tokens, or the model's shorter maximum, plus 0.8 GB for the software. The total is an estimate.
| Rank | AI model | Parameters | Best quant | GB | Label |
|---|---|---|---|---|---|
| 1 | Can I run DeepSeek V4.1 Flash 1017? | 763.21B (estimate), 16B active | UD-IQ1_S, Engram table on SSD | 123.18 GB | Unsloth guide: 1-bit needs 128–145 GB RAM+VRAM with the Engram table on SSD via mmap; not tested by us. Our estimate leaves out the conversation memory for this model: real use needs more than the figures shown |
| 2 | Can I run Reflection Beam 1.0? | 501.3B, 23.4B active | UD-IQ1_M | 124.80 GB | measured weights; total is an estimate |
| 3 | Can I run Qwen3.8-Flash-Next? | 180B (estimate), 6B active | UD-Q4_K_XL | 112.13 GB | measured weights; total is an estimate. Our estimate leaves out the conversation memory for this model: real use needs more than the figures shown |
| 4 | Can I run Ornith-1.5-122B? | 125.09B | UD-Q6_K_XL | 115.22 GB | measured weights; total is an estimate |
| 5 | Can I run NVIDIA Nemotron 3.5 Super VL? | 120B (estimate), 12B active | Q5 (~0.71 bytes/param) | 86.07 GB | estimate |
| 6 | Can I run Laguna S 2.1? | 118B (estimate), 8B active | UD-Q6_K_XL | 108.36 GB | measured weights; total is an estimate |
| 7 | Can I run gpt-oss-120b? | 117B (estimate), 5.1B active | F16 | 66.48 GB | measured weights; total is an estimate |
| 8 | Can I run Llama 4 Scout? | 108.64B (estimate), 17B active | Q8_0 | 116.94 GB | measured weights; total is an estimate |
| 9 | Can I run Kolibri-1? | 78.1B (estimate), 3.46B active | Q8_0 | 84.15 GB | measured weights; total is an estimate |
| 10 | Can I run NVIDIA Nemotron 3.5 Bolt 74B-A9B? | 74B (estimate), 9B active | Q8 | 80.05 GB | estimate |
| 11 | Can I run Qwen2.5 72B? | 72.7B (estimate) | Q8 | 81.27 GB | estimate |
| 12 | Can I run DeepSeek-R1-Distill 70B? | 70.6B (estimate) | Q8 | 79.03 GB | estimate |
| 13 | Can I run Llama 3.3 70B? | 70.6B (estimate) | Q8 | 79.03 GB | estimate |
| 14 | Can I run Recursive Flint? | 39.79B, 2.5B active | BF16 | 80.70 GB | measured weights; total is an estimate |
| 15 | Can I run Qwen3.6-35B-A3B? | 35.95B (estimate), 3B active | FP16 | 72.87 GB | estimate |
| 16 | Can I run Nanbeige4.3-33B? | 33.12B (estimate) | BF16 | 68.78 GB | measured weights; total is an estimate |
| 17 | Can I run Qwen2.5 32B? | 32.8B (estimate) | FP16 | 68.55 GB | estimate |
| 18 | Can I run Qwen3 32B? | 32.76B (estimate) | FP16 | 68.47 GB | estimate |
| 19 | Can I run Gemma 4 31B? | 31.27B (estimate) | BF16 | 63.72 GB | measured weights; total is an estimate |
| 20 | Can I run Qwen3 Coder 30B A3B? | 30.53B (estimate), 3.3B active | FP16 | 62.67 GB | estimate |
| 21 | Can I run NVIDIA Nemotron 3.5 Lightning? | 30B (estimate), 3B active | BF16 | 64.03 GB | measured weights; total is an estimate |
| 22 | Can I run Muse Glimmer 30B? | 29.6B | BF16 | 60.54 GB | safetensors listing, not a GGUF |
| 23 | Can I run Qwen3.8-27B? | 27.78B (estimate) | FP16 | 56.90 GB | estimate |
| 24 | Can I run Gemma 3 27B? | 27.43B (estimate) | F16 | 55.94 GB | measured weights; total is an estimate |
| 25 | Can I run Gemma 2 27B? | 27.2B (estimate) | FP16 | 57.52 GB | estimate |
| 26 | Can I run OrcaCyber Nano 1.0? | 27B | FP16 | 55.34 GB | estimate |
| 27 | Can I run Gemma 4 26B A4B? | 25.2B (estimate), 3.8B active | BF16 | 51.69 GB | measured weights; total is an estimate |
| 28 | Can I run Mistral Small 3.2 24B? | 24.01B (estimate) | BF16 | 49.29 GB | measured weights; total is an estimate |
| 29 | Can I run gpt-oss-20b? | 21B (estimate), 3.6B active | F16 | 14.79 GB | measured weights; total is an estimate |
| 30 | Can I run DeepSeek R1 Distill Qwen 14B? | 14.77B | F16 | 31.96 GB | measured weights; total is an estimate |
| 31 | Can I run Phi-4 14B? | 14.7B (estimate) | FP16 | 31.88 GB | estimate |
| 32 | Can I run Qwen2.5 14B? | 14.7B (estimate) | FP16 | 31.81 GB | estimate |
| 33 | Can I run Gemma 2 9B? | 9.24B (estimate) | FP16 | 21.39 GB | estimate |
| 34 | Can I run Llama 3.1 8B? | 8.03B (estimate) | FP16 | 17.93 GB | estimate |
| 35 | Can I run NVIDIA Nemotron 3.5 Bolt E4B? | 8B (estimate) | BF16 | 16.12 GB | measured weights; total is an estimate |
| 36 | Can I run Mistral 7B v0.3? | 7.25B (estimate) | FP16 | 16.37 GB | estimate |
| 37 | Can I run Spark-X2.6-4B? | 4.11B (estimate) | BF16 | 9.39 GB | measured weights; total is an estimate |
| 38 | Can I run Llama 3.2 3B? | 3.21B (estimate) | FP16 | 8.16 GB | estimate |
| 39 | Can I run LFM3-1.7B? | 1.73B (estimate) | BF16 | 4.38 GB | measured weights; total is an estimate |
Cards and Macs with 128 GB
The euro price is the figure we store, with the date it was written down. A Mac's gigabytes are shared memory, not graphics memory, so a Mac does not hold the same models as a 128 GB card.
| Machine | Memory | EU price | Date | Label |
|---|---|---|---|---|
| Mac Studio M4 Max 128GB | 128 GB shared memory | €3,290 | price unverified | |
| Mac Studio M5 Max 128GB | 128 GB shared memory | €6,000 | price unverified |
Memory totals are our estimate, including when the weight file was measured. Prices are the euro figures we store, with the date they were written down. How we calculate
Common questions
- What is the largest model that fits in 128 GB of VRAM?
- DeepSeek V4.1 Flash 1017 is the largest by parameter count, at UD-IQ1_S, Engram table on SSD, 123.18 GB (Unsloth guide: 1-bit needs 128–145 GB RAM+VRAM with the Engram table on SSD via mmap; not tested by us. Our estimate leaves out the conversation memory for this model: real use needs more than the figures shown). 39 models on our list fit.
- What does the best quant mean for 128 GB?
- The best quant is the highest-quality option that still fits. A measured download is used when we have one at that quality. Otherwise the figure is Q4, Q5, Q8 or FP16 from our estimate. A higher bit count wins. A file that runs only on a Mac is not counted. The gigabytes add conversation memory and the software overhead, so the total is an estimate even when the weight file was measured.
- Which cards and Macs have 128 GB, and what do they cost in the EU?
- Mac Studio M4 Max 128GB (128 GB shared memory): €3,290, price unverified, 16 Oct 2026. Mac Studio M5 Max 128GB (128 GB shared memory): €6,000, price unverified, 16 Oct 2026.
- How much of 128 GB can a model use?
- Our estimate keeps 0.6 GB free on one graphics card, so 127.4 GB is usable. A Mac with 128 GB of shared memory does not give the graphics all of it. The model table is for one graphics card.