Overview
Kimi K3 VRAM requirements are far beyond any single consumer GPU: the model card lists 2.8T total parameters, and even at a bare 4 bits per weight that is 2.8e12 × 4 ÷ 8 ÷ 1e9 = 1,400 decimal GB for weights alone, before KV cache and overhead. The card states no memory requirement, so every size below is arithmetic from the card’s parameter count, not a published file size. Two of the six models here (Nemotron 3.5 Lightning and Granite 4.2) are small enough for one workstation GPU; the rest are multi-GPU or cluster models. Tokens per second are unknown on this page, and prices: check current retailer price.
How are these sizes worked out?
The formula is from methodology: weights_GB = params × bits ÷ 8 ÷ 1e9, plus KV cache, plus a labelled overhead (an assumption of 1 to 2 GB on this site, not a measurement). No config.json was read for this page, so the KV cache term is unknown for every model here. A mixture-of-experts model keeps all experts in memory, so total parameters set the weight size and active parameters only change per-token work. See moe-cpu-offload for what offloading experts means.
| Model | Total params (card) | Active | 16-bit weights | 8-bit weights | 4-bit floor |
|---|---|---|---|---|---|
| Kimi K3 | 2.8T | 104B | 5,600 GB | 2,800 GB | 1,400 GB |
| GLM-5.2 | 753B (metadata) | unknown | 1,506 GB | 753 GB | 376.5 GB |
| MiniMax M3 | ~428B | ~23B | 856 GB | 428 GB | 214 GB |
| GLM-5.3-Flash | 320B | 18B | 640 GB | 320 GB | 160 GB |
| Nemotron 3.5 Lightning | 30B | 3B | 60 GB | 30 GB | 15 GB |
| Granite 4.2 30B | 30B | not stated (dense) | 60 GB | 30 GB | 15 GB |
The 4-bit column uses exactly 4.000 bits and ignores quantization scales and any tensors kept at higher precision, so real 4-bit files are larger. These are decimal GB of weights only. The cards’ own metadata counts differ slightly (for example 32B for Nemotron, 29B for Granite, 427B for MiniMax M3, 321B for GLM-5.3-Flash), so treat each figure as a rough size. For your own numbers use vram-calculator when it ships and vram-requirements for the dense worked examples.
Can I run Kimi K3 at home?
Not on a gaming or single-workstation GPU. The Kimi K3 card lists 2.8T total and 104B activated parameters, a 1,048,576-token context length, 93 layers and 896 experts with 16 selected per token, and “MXFP4 weights / MXFP8 activations” from quantization-aware training. The weights are already near 4-bit, so the 1,400 GB floor above is close to the real scale. For comparison, an RTX 5090 has “32 GB GDDR7” per NVIDIA’s page, and 1,400 ÷ 32 = 43.75 cards for weights alone. The card recommends vLLM, SGLang and TokenSpeed and gives no VRAM or GPU requirement. The card’s repo size is not stated, and so is the KV cache. For a smaller MoE example that is still big, see deepseek-vram-requirements; for large unified-memory boxes see dgx-spark-vs-strix-halo.
How much VRAM does GLM 5.3 Flash need?
At least 160 GB at the 4-bit floor, and 320 GB at 8 bits. The GLM-5.3-Flash card says “320B total parameters” and “18B active parameters” under an MIT license, with tensor types BF16, F8_E4M3 and F32 (the metadata shows 321B). Because the card mixes precisions, the repo size is unknown here (the card does not state it). The card gives no context specification and no memory requirement; its 1M, 400K and 164K context figures are benchmark evaluation settings, not specifications. Against a 32 GB RTX 5090 the weights are 160 ÷ 32 = 5 cards at the very least, so a single 5090 does not hold it without heavy expert offload to system RAM (ram-vs-vram, moe-cpu-offload).
How much VRAM does GLM 5.2 need?
Unknown from the card text. The GLM-5.2 card prose does not state total or active parameters; only the Hugging Face metadata shows 753B, with tensor types BF16 and F32 and an MIT license and “1M-token context”. If the weights were stored as BF16 the arithmetic is 753e9 × 16 ÷ 8 ÷ 1e9 = 1,506 GB; the real repo size is not stated. The “4 CPUs, 8 GB RAM” on the card describes benchmark sandboxes, not serving requirements.
How much memory does MiniMax M3 need?
About 214 GB at a 4-bit floor, 428 GB at 8 bits, and 856 GB at BF16, from the card’s “~428B parameters” and “~23B activated parameters”. The MiniMax-M3 card lists “1M context”, tensor types BF16 and F32, and the minimax-community license. It gives no memory requirement or repo size. The listing on Ollama shows 522.7K downloads (a download counter, not a search volume) and a 512K context window on its one visible tag, minimax-m3:cloud.
Which Ollama tags are cloud-only and which are local?
For the two models checked, the only tag visible on 2026-10-13 was a :cloud tag. The kimi-k3 library page and its tags page list only kimi-k3:cloud (1M context, size not stated); the minimax-m3 page lists only minimax-m3:cloud. Ollama’s cloud docs say cloud models “do not need to be downloaded” and that “Ollama processes cloud prompts and responses”, so a :cloud tag is not running on your GPU. Whether every other model has a local tag was not verified: unknown.
Can Nemotron 3.5 Lightning and Granite 4.2 run on one GPU?
Yes for Nemotron on an 80 GB card, and plausibly for both at quantized sizes. The Nemotron 3.5 Lightning card lists 30B total and 3B active, BF16 weights, and says its context is “Up to 1M tokens (for single H100 deployment, we use 256K)”, and that BF16 on one H100 is memory-bound to about 256K (the full 1M window was validated on 8x H100, 1x GB200 or 1x B200). BF16 weights are 30e9 × 16 ÷ 8 ÷ 1e9 = 60 GB. The Granite 4.2 30B card lists a dense 30B (variants 3B, 8B, 30B), BF16, Apache 2.0 and 128K native context (512K with extension); 60 GB at BF16, and the card gives no VRAM requirement. At the 4-bit floor both are 15 GB, which is why a 16 GB card is tight rather than comfortable once KV cache and overhead are added. No 16 GB claim is made on either card: unknown. See 48gb-vram-llm for a 48 GB class.
Can I run DeepSeek or Qwen locally?
Sizes depend on the variant. DeepSeek V4 Flash is a 160 GB repo and the dense Qwen and R1 distills are far smaller; both are worked out on deepseek-vram-requirements and qwen-vram-requirements. Mistral Medium 3.5 128B and the older Kimi K2 and GLM 4.5 Air were not read for this page: unknown.
Questions
Can I run Kimi K3 at home?
Not on a single consumer GPU. The Kimi K3 card lists 2.8T total parameters stored as MXFP4, so weights are at least 2.8e12 × 4 ÷ 8 ÷ 1e9 = 1,400 GB, about 44 RTX 5090 cards of 32 GB each. The card states no hardware requirement, and the only Ollama tag seen is kimi-k3:cloud, which runs on Ollama’s servers.
How much VRAM does GLM 5.3 Flash need?
At least 160 GB for weights at a 4-bit floor, and 320 GB at 8 bits, from 320B total parameters on its card. The card does not state a memory requirement, repo size or default context, so KV cache and the exact size are unknown. A single 32 GB RTX 5090 cannot hold the weights without offloading experts to system RAM.
How much memory does MiniMax M3 need?
About 214 GB at a 4-bit floor, 428 GB at 8 bits and 856 GB at BF16, from “~428B parameters” on the card. The card gives no memory requirement or repo size, and the formula leaves out KV cache for its 1M context, which is unknown.
Which Ollama tags are cloud-only vs local?
For kimi-k3 and minimax-m3, the only tag visible on 2026-10-13 was a :cloud tag. Ollama’s docs say cloud models need no download and are processed by Ollama, so they do not use your GPU. Whether other models in the library have local tags was not checked: unknown.
Can I run DeepSeek or Qwen locally?
Yes for smaller variants, no for the largest. Dense Qwen and DeepSeek R1 distill sizes can be computed with the standard formula on methodology; the full DeepSeek V4 Flash needs about 160 GB. See deepseek-vram-requirements and qwen-vram-requirements.
Do these tables include speed?
No. Tokens per second are unknown because none of the linked cards publish a figure with a named machine.
Sources read on 2026-10-13
- Kimi K3 model card: 2.8T total, 104B activated, 1,048,576 context, MXFP4/MXFP8, Kimi K3 License, 93 layers, 896 experts, 16 per token, vLLM/SGLang/TokenSpeed; no memory requirement.
- GLM-5.3-Flash card: 320B total (321B metadata), 18B active, MIT, BF16/F8_E4M3/F32; no context spec, memory or repo size.
- GLM-5.2 card: 753B in metadata only, active not stated, 1M-token context, MIT, BF16/F32.
- MiniMax-M3 card: ~428B (427B metadata), ~23B activated, 1M context, minimax-community license, BF16/F32.
- Nemotron 3.5 Lightning 30B-A3B BF16 card: 30B/3B active, BF16, 256K fits one 80 GB H100, openmdw-1.1.
- Granite 4.2 30B card: 30B (3B/8B/30B variants), BF16, Apache 2.0, 128K native, no memory requirement.
- Ollama kimi-k3, tags, minimax-m3: only
:cloudtags visible; 522.7K downloads for minimax-m3. - Ollama cloud docs: cloud models need no download and are processed by Ollama.
- NVIDIA RTX 5090: 32 GB GDDR7.
Related
- methodology is the formula and citation rule.
- vram-requirements and vram-calculator size dense models.
- deepseek-vram-requirements and qwen-vram-requirements cover two other families.
- moe-cpu-offload and ram-vs-vram cover running a large MoE across RAM and VRAM.
- dgx-spark-vs-strix-halo, 48gb-vram-llm and multi-gpu-llm cover larger-memory hardware.