The biggest open models are free to download and impossible to fit on a gaming PC. How much memory DeepSeek V4.1 Flash, Kimi K3, GLM-5.3, MiniMax, gpt-oss-120b, Llama 4 and the 70B models really need, the cheapest box that holds each, and when an H100, H200 or B200 by the hour beats buying.
Some links on this page are referral links: we may earn a commission, at no cost to you.
Share
Giant open models need server memory, not a gaming GPU. On our VRAM engine a 70B model needs about 46 GB at Q4_K_M, gpt-oss-120b about 63 GB, MiniMax M2.5 about 134 GB, GLM-5.3 about 427 GB, DeepSeek V4.1 Flash about 476 GB and Kimi K3 about 1,455 GB. Up to about 120B parameters you can own the hardware (a 48 to 96GB card, two used 3090s or a big-memory Mac); above that, renting H100, H200 or B200 servers by the hour, or paying per token for the API, is the only sensible way for almost everyone.
The memory column is our VRAM engine's estimate at 16K tokens of context (the number on each model's cost-to-run page, read Nov 17, 2026), at Q4_K_M or the model's own low-bit release. Download sizes are from the linked guides. "Fits" counts GPU memory only; MoE models can also run slower with experts in system RAM.
Llama 3.3 70B
Total / active parameters
70B dense
Memory needed
about 46 GB (Q4_K_M)
Smallest setup that holds it
One 48GB card (RTX A6000) or two 24GB cards
DeepSeek R1 70B distill
Total / active parameters
70B dense
Memory needed
43 GB download
Smallest setup that holds it
One 48GB card or two 24GB cards
gpt-oss-120b
Total / active parameters
117B / 5.1B active
Memory needed
about 63 GB (MXFP4)
Smallest setup that holds it
One 80GB H100 or A100, or a 96GB card
Llama 4 Scout
Total / active parameters
109B / 17B active
Memory needed
about 63 GB (Q4_K_M)
Smallest setup that holds it
One 80GB H100
GLM-4.5-Air
Total / active parameters
106B / 12B active
Memory needed
about 60 GB (4-bit)
Smallest setup that holds it
One 80GB card or a 128GB unified-memory box
MiniMax M2.5
Total / active parameters
228.7B MoE
Memory needed
about 134 GB (Q4_K_M)
Smallest setup that holds it
Two 80GB H100s or one B200
MiniMax M3.1 Flash
Total / active parameters
427.1B MoE
Memory needed
236 GB (its NVFP4 release)
Smallest setup that holds it
Two H200s or two B200s
Llama 4 Maverick
Total / active parameters
400B MoE
Memory needed
about 243 GB
Smallest setup that holds it
Four H100s or two H200s
GLM-5.3
Total / active parameters
753.3B MoE
Memory needed
about 427 GB (Q4_K_M)
Smallest setup that holds it
A four-GPU H200 or B200 server
DeepSeek V4.1 Flash
Total / active parameters
763B incl. Engram table
Memory needed
about 476 GB (its 4-bit release)
Smallest setup that holds it
Four H200s, four B200s or eight H100s
Kimi K3
Total / active parameters
2.8T / 104B active
Memory needed
about 1,455 GB (MXFP4)
Smallest setup that holds it
A multi-node cluster (vLLM asks for 8+ GB300s)
Total / active parameters
Memory needed
Smallest setup that holds it
Llama 3.3 70B
70B dense
about 46 GB (Q4_K_M)
One 48GB card (RTX A6000) or two 24GB cards
DeepSeek R1 70B distill
70B dense
43 GB download
One 48GB card or two 24GB cards
gpt-oss-120b
117B / 5.1B active
about 63 GB (MXFP4)
One 80GB H100 or A100, or a 96GB card
Llama 4 Scout
109B / 17B active
about 63 GB (Q4_K_M)
One 80GB H100
GLM-4.5-Air
106B / 12B active
about 60 GB (4-bit)
One 80GB card or a 128GB unified-memory box
MiniMax M2.5
228.7B MoE
about 134 GB (Q4_K_M)
Two 80GB H100s or one B200
MiniMax M3.1 Flash
427.1B MoE
236 GB (its NVFP4 release)
Two H200s or two B200s
Llama 4 Maverick
400B MoE
about 243 GB
Four H100s or two H200s
GLM-5.3
753.3B MoE
about 427 GB (Q4_K_M)
A four-GPU H200 or B200 server
DeepSeek V4.1 Flash
763B incl. Engram table
about 476 GB (its 4-bit release)
Four H200s, four B200s or eight H100s
Kimi K3
2.8T / 104B active
about 1,455 GB (MXFP4)
A multi-node cluster (vLLM asks for 8+ GB300s)
The pattern that catches people: "A3B", "17B active" or "104B active" is how much of the model works per token, which sets speed. Memory is set by total parameters, because every expert must be stored somewhere. That's why Kimi K3 needs about 1.56 TB for its official MXFP4 weights even though only about 104B parameters run per token.
Which giant model can you actually run, and on what?
70B models on 48GB
A 70B dense model like Llama 3.3 70B needs about 42.5 GB at Q4_K_M before context, so you need roughly 48 GB of GPU or unified memory. Three ways to get there, priced Nov 3, 2026 in the cheapest way to run a 70B model:
Rent a 48GB card such as an RTX A6000. It's the cheapest for light or occasional use.
Two used RTX 3090s, about $1,000 per card, the cheapest hardware to own. They break even against renting after roughly 3,400 to 6,100 hours, depending on the rental rate.
A Mac Studio M5 Max with 64GB, from $3,499: quiet and one box, but the most expensive of the three.
DeepSeek R1's 70B distill is a 43 GB Ollama download and fits the same setups; the smaller distills run on gaming cards (8B on 8GB, 14B on 16GB, 32B on 24GB). See which DeepSeek R1 model fits your GPU.
gpt-oss-120b, Llama 4 Scout and GLM-4.5-Air on 64 to 96GB
gpt-oss-120b needs about 61 GiB for its weights and about 64 GiB in total with 32K context. Every GGUF from Q2_K to Q8_0 is 62.6 to 63.4 GB, because its experts stay in OpenAI's MXFP4 format, so a smaller quant won't save you. One 80GB H100 or A100, a 96GB RTX PRO 6000 Blackwell, or a 96 to 128GB unified-memory machine holds it. A 24GB card also works with 64GB of system RAM and `--n-cpu-moe` expert offload, more slowly.
Llama 4 Scout (109B, 17B active) is about 62 GB at Q4_K_M, so no single consumer GPU holds a good quant; a very low 1.78-bit build of about 33 GB fits a 32GB card. Maverick (400B) isn't a home model. See how to run Llama 4 locally. GLM-4.5-Air (106B, 12B active) is about 60 GB at 4-bit, the one GLM that fits a single 80GB card or a 128GB unified box, per how to run GLM locally.
MiniMax, GLM-5.3 and DeepSeek V4.1 Flash, server territory
These are where buying stops making sense for one person:
MiniMax M2.5, 228.7B parameters with 8 of 256 experts active, needs about 134 GB at Q4_K_M on our engine. See what it costs to run MiniMax M2.5.
MiniMax M3.1 Flash, 427.1B parameters, ships in NVFP4 and needs about 236 GB in that form. NVFP4 runs at full speed only on Blackwell GPUs such as the B200; see what NVFP4 is and the GPUs that run MiniMax M3.1 Flash.
GLM-5.3, 753.3B parameters, needs about 427 GB at Q4_K_M on our engine; its own FP8 release is a 755.6 GB download. See GLM-5.3's costs.
DeepSeek V4.1 Flash counts 763B parameters on Hugging Face, a 552B backbone plus a 196B Engram memory table, with 16B active while it writes. DeepSeek's own release is already 4-bit, a 510.3 GB download that needs about 476 GB. Four H200s, four B200s or eight H100s hold it, or tightly a Mac Studio with 512GB of unified memory. Unsloth's 1-bit build runs on 128 to 145 GB of RAM plus VRAM with the Engram table on an SSD, at lower accuracy. See can you run DeepSeek V4.1 Flash locally.
Kimi K3 and anything above 1TB
Kimi K3 has 2.8 trillion parameters with about 104B active per token. Moonshot's official MXFP4 weights total about 1.56 TB and the smallest community GGUF is 509 GB; vLLM's recipe asks for at least 8 GB300 GPUs. No home rig and no single eight-GPU B200 box (1,440 GB in total) holds the official weights. For nearly everyone the answer is the API: Moonshot charges $3.00 per 1M input tokens and $15.00 per 1M output tokens for kimi-k3 (re-checked Nov 12, 2026). See what it takes to run Kimi K3 and how to get a Kimi K3 API key.
When does renting beat buying for a giant model?
Three honest rules, using our price feed (on-demand, one GPU, Nov 17, 2026):
Under 80 GB, renting is cheap to try. On Vast.ai an RTX A6000 (48GB) started at $0.32 an hour and an H100 SXM (80GB) at $1.53; Runpod listed them at $0.69 and $4.39. A weekend of tests costs less than one used 3090. Hosts compared in the cheapest H100 rental.
From 140 GB to 500 GB, renting is the only realistic way to run it yourself. An H200 (141GB) started at $3.75 an hour on Vast.ai and $5.29 on Runpod; a B200 at $7.50 and $7.99. Four H200s on Runpod come to $21.16 an hour. See what an H200 costs to rent.
For chat, the API usually wins. DeepSeek's API charges $0.30 per 1M input tokens and $1.20 per 1M output tokens at peak, half off-peak (read Nov 7, 2026). At $21.16 an hour, a four-H200 server only beats it if it turns out more than about 17.6 million output tokens every hour. Rent the server for private data, fine-tuning, batch jobs or a product that keeps it busy.
Not the full models. On our engine DeepSeek V4.1 Flash needs about 476 GB, GLM-5.3 about 427 GB at Q4_K_M and Kimi K3 about 1,455 GB. What runs at home is DeepSeek's distilled R1 models (8B to 32B on 8 to 24GB cards) or GLM-4.5-Air on an 80GB card or a 128GB unified-memory box.
How much VRAM do I need for a 70B model?
About 42.5 GB for the weights at Q4_K_M and about 46 GB with 16K context, so a 48GB card such as an RTX A6000, two 24GB cards, or a Mac with 64GB of unified memory.
Does a mixture-of-experts model need less memory?
No. Every expert has to be loaded, so memory follows total parameters. Active parameters set speed. That's why Kimi K3, with about 104B active, still needs about 1.56 TB for its official weights.
What GPU do I rent for gpt-oss-120b?
One 80GB H100 or A100, since it needs about 64 GiB with 32K context. On Nov 17, 2026 an H100 SXM started at $1.53 an hour on Vast.ai and $4.39 on Runpod.
Is renting H200s cheaper than the DeepSeek API?
Only for heavy, steady use. Four H200s on Runpod cost $21.16 an hour (Nov 17, 2026), which beats DeepSeek's $1.20 per 1M output tokens only above about 17.6 million output tokens an hour. For chat and light use, the API is cheaper.
Can a Mac run these giant models?
Some of them. A 64GB Mac runs a 70B model at 4-bit, a 128GB machine runs gpt-oss-120b or GLM-4.5-Air, and a Mac Studio with 512GB of unified memory holds DeepSeek V4.1 Flash at 4-bit, tightly. Kimi K3 doesn't fit any single machine.