# LLMHardware.io

> Hardware guidance for running open-source LLMs locally. Kept guides, a hardware index, a compare tool, and a VRAM calculator. There are no per-model or per-SKU detail pages.

LLMHardware.io is an independent reference site. It is NOT affiliated with Hugging Face, NVIDIA, AMD, or Apple. VRAM figures use the published formula below. Always check a current retailer price before buying.

## What this site publishes

- 57 kept guides (GPU buying, Apple Silicon, setup, quantization, model-family requirements, and local alternatives to cloud agents)
- One hardware index with specs for the GPUs and Apple Silicon devices we track
- A compare tool and a VRAM calculator
- Trust pages (about, methodology, editorial policy, disclosures, privacy, terms)

It does not publish a page per Hugging Face model, a page per GPU SKU, or embed-only URLs.

## Key tools

- VRAM Calculator: https://llmhardware.io/vram-calculator
- Can I Run It? (GPU/Mac model fit checker): https://llmhardware.io/can-i-run
- Hardware index: https://llmhardware.io/hardware
- Compare: https://llmhardware.io/compare
- Guides: https://llmhardware.io/guides
- Full text for language models (page digests and FAQs): https://llmhardware.io/llms-full.txt

## Trust and about

- Methodology: https://llmhardware.io/methodology
- About: https://llmhardware.io/about
- Author: https://llmhardware.io/authors/llm-hardware-team
- Contact: https://llmhardware.io/contact
- Editorial policy: https://llmhardware.io/editorial-policy
- Disclosures: https://llmhardware.io/disclosures
- Privacy: https://llmhardware.io/privacy
- Terms: https://llmhardware.io/terms

## Methodology summary

VRAM requirement formula:

    VRAM (GB) = parameters(B) × bytes_per_param × 1.2 + 2 GB OS overhead

Quantization bytes_per_param values: FP16 = 2.0 | Q8 = 1.0 | Q5_K_M = 0.625 | Q4_K_M = 0.5

Worked short-context checks, 2026-10-09:

- Qwen3.8-27B at Q4_K_M: 27 × 0.5 × 1.2 + 2 = 18.2 GB. Unsloth UD-Q4 files that day were 15.4–17.6 GB. Ollama tag qwen3.8:27b listed 18 GB.
- Gemma 4 26B-A4B at Q4_K_M, using 25.2B total (not 3.8B active): 17.1 GB. Unsloth’s 4-bit note is 16–18 GB.
- Aleph Alpha Kolibri-1 is about 78 GB FP8 on its model card and is not a 24 GB consumer model.
- Qwen3.8-Flash-Next, counting 125B + 51B n-gram + 4B MTP = 180B: 180 × 0.6 + 2 = 110 GB. Unsloth UD-Q4_K_XL summed to 111.33 GB on 2026-10-09. A 32 GB GPU does not fit it.

For MoE models, put total parameters in the formula. Active parameters change speed, not the weight footprint.
Long context adds KV cache on top of the formula: 2 × layers × KV heads × head dim × bytes per token. Qwen3-32B at a 32K window is about 8.6 GB at the f16 default and about half at q8_0 (OLLAMA_KV_CACHE_TYPE, Ollama FAQ). Details are in the How Much VRAM guide (KV cache section).
MoE expert weights can be kept in system RAM with llama.cpp --cpu-moe / --n-cpu-moe N, slower than a full GPU fit.

<!-- generated:tables:start (scripts/generate-llms.mjs) -->

## Model memory table (32 local models)

Formula estimates at a short context (about 4K tokens): params × bytes × 1.2 + 2 GB. Q4_K_M = 0.5 bytes, Q8_0 = 1 byte. MoE models use total parameters. Figures of 100 GB and up are rounded to the nearest GB. Published GGUF files are usually smaller than these estimates.

| Model | Params (B) | Q4_K_M (GB) | Q8 (GB) | Page |
|---|---|---|---|---|
| Dory 1.0 4B | 3.98 | 4.4 | 6.8 | https://llmhardware.io/guides/best-llm-for-coding |
| Qwen3 4B | 4 | 4.4 | 6.8 | https://llmhardware.io/guides/qwen3-hardware-requirements |
| Mistral 7B | 7 | 6.2 | 10.4 | https://llmhardware.io/guides/mistral-hardware-requirements |
| Qwen3 8B | 8 | 6.8 | 11.6 | https://llmhardware.io/guides/qwen3-hardware-requirements |
| Llama 3.1 8B | 8 | 6.8 | 11.6 | https://llmhardware.io/guides/llama-3.1-hardware-requirements |
| Granite 4.2 8B | 8 | 6.8 | 11.6 | https://llmhardware.io/vram-calculator |
| Granite Switch 4.1 8B | 8.78 | 7.3 | 12.5 | https://llmhardware.io/vram-calculator |
| LFM2.5-8B-A1B (MoE) | 8.3 | 7.0 | 12.0 | https://llmhardware.io/guides/lfm2-5-8b-a1b-hardware-requirements |
| Gemma 3 12B | 12 | 9.2 | 16.4 | https://llmhardware.io/guides/gemma3-hardware-requirements |
| Mellum2.1 12B-A2.5B (MoE) | 12 | 9.2 | 16.4 | https://llmhardware.io/vram-calculator |
| Qwen3 14B | 14 | 10.4 | 18.8 | https://llmhardware.io/guides/qwen3-hardware-requirements |
| Phi-4 14B | 14 | 10.4 | 18.8 | https://llmhardware.io/guides/phi4-hardware-requirements |
| Mistral Small 3.1 24B | 24 | 16.4 | 30.8 | https://llmhardware.io/guides/mistral-hardware-requirements |
| LFM2.5-24B-A2B (MoE) | 24 | 16.4 | 30.8 | https://llmhardware.io/guides/lfm2-5-8b-a1b-hardware-requirements |
| Gemma 4 26B-A4B (MoE, 25.2B total) | 25.2 | 17.1 | 32.2 | https://llmhardware.io/vram-calculator |
| Gemma 3 27B | 27 | 18.2 | 34.4 | https://llmhardware.io/guides/gemma3-hardware-requirements |
| Qwen3.6-27B | 27 | 18.2 | 34.4 | https://llmhardware.io/vram-calculator |
| Qwen3.8-27B | 27 | 18.2 | 34.4 | https://llmhardware.io/guides/local-models-2026-hardware |
| Qwen3 30B-A3B (MoE) | 30 | 20.0 | 38.0 | https://llmhardware.io/guides/qwen3-hardware-requirements |
| Granite Switch 4.1 30B | 29.99 | 20.0 | 38.0 | https://llmhardware.io/guides/local-models-2026-hardware |
| Granite 4.2 30B | 30 | 20.0 | 38.0 | https://llmhardware.io/vram-calculator |
| Qwen3 32B | 32 | 21.2 | 40.4 | https://llmhardware.io/guides/qwen3-hardware-requirements |
| DeepSeek R1 Distill Qwen 32B | 32 | 21.2 | 40.4 | https://llmhardware.io/guides/deepseek-hardware-requirements |
| Qwen3.6-35B-A3B (MoE) | 35 | 23.0 | 44.0 | https://llmhardware.io/vram-calculator |
| Llama 3.3 70B | 70 | 44.0 | 86.0 | https://llmhardware.io/guides/llama-3.1-hardware-requirements |
| Nemotron 3.5 Bolt 74B-A9B (MoE) | 74 | 46.4 | 90.8 | https://llmhardware.io/guides/nemotron-3-5-bolt-hardware-requirements |
| Llama 4 Scout 17B-16E (MoE, 109B total) | 109 | 67.4 | 133 | https://llmhardware.io/guides/llama-3.1-hardware-requirements |
| Nemotron 3.5 Super VL 120B-A12B (MoE) | 120 | 74.0 | 146 | https://llmhardware.io/guides/best-llms-for-64gb-128gb-ram |
| Qwen3 235B-A22B (MoE) | 235 | 143 | 284 | https://llmhardware.io/guides/qwen3-hardware-requirements |
| Beam 501B-A23B (MoE) | 501.3 | 303 | 604 | https://llmhardware.io/guides/best-llms-for-64gb-128gb-ram |
| DeepSeek R1 671B (MoE) | 671 | 405 | 807 | https://llmhardware.io/guides/deepseek-hardware-requirements |
| DeepSeek-V4.1-Flash-1017 (MoE) | 763.2 | 460 | 918 | https://llmhardware.io/guides/deepseek-hardware-requirements |

## Hardware table (31 devices)

VRAM or unified memory and memory bandwidth from vendor spec sheets. Apple unified memory is shared with macOS: subtract about 4 GB before comparing with the table above. No prices: check the live price at the retailer.

| Device | Memory (GB) | Type | Bandwidth (GB/s) |
|---|---|---|---|
| NVIDIA GeForce RTX 4060 (8GB) | 8 | VRAM | 272 |
| Intel Arc B580 (12GB) | 12 | VRAM | 456 |
| NVIDIA GeForce RTX 4070 (12GB) | 12 | VRAM | 504 |
| NVIDIA RTX 5070 (12GB) | 12 | VRAM | 672 |
| Apple Mac mini (M4, 16GB) | 16 | unified | 120 |
| Apple Mac mini (M6, 16GB) | 16 | unified | 153 |
| NVIDIA RTX 4060 Ti (16GB) | 16 | VRAM | 288 |
| NVIDIA RTX 4070 Ti Super (16GB) | 16 | VRAM | 672 |
| NVIDIA RTX 4080 (16GB) | 16 | VRAM | 716 |
| NVIDIA RTX 5070 Ti (16GB) | 16 | VRAM | 896 |
| NVIDIA RTX 5080 (16GB) | 16 | VRAM | 960 |
| PNY GeForce RTX 5060 Ti (16GB) | 16 | VRAM | 448 |
| AMD Radeon RX 7900 XTX (24GB) | 24 | VRAM | 960 |
| Apple Mac mini (M4 Pro, 24GB) | 24 | unified | 273 |
| Apple Mac mini (M4, 24GB) | 24 | unified | 120 |
| Apple Mac mini (M5 Pro, 24GB) | 24 | unified | 307 |
| Apple Mac mini (M6, 24GB) | 24 | unified | 170 |
| NVIDIA GeForce RTX 3090 (24GB) | 24 | VRAM | 936 |
| NVIDIA GeForce RTX 4090 (24GB) | 24 | VRAM | 1008 |
| NVIDIA GeForce RTX 5090 (32GB) | 32 | VRAM | 1792 |
| Apple Mac Studio (M5 Max, 36GB) | 36 | unified | 460 |
| Apple Mac mini (M4 Pro, 48GB) | 48 | unified | 273 |
| Apple Mac Studio (M5 Max, 48GB) | 48 | unified | 614 |
| Apple Mac Studio (M4 Max, 64GB) | 64 | unified | 546 |
| Apple Mac Studio (M5 Max, 64GB) | 64 | unified | 614 |
| Apple Mac Studio (M5 Ultra, 96GB) | 96 | unified | 1200 |
| Apple Mac Studio (M4 Max, 128GB) | 128 | unified | 546 |
| Apple Mac Studio (M5 Max, 128GB) | 128 | unified | 614 |
| Beelink GTR9 Pro (128GB) | 128 | unified | 256 |
| GMKtec EVO-X2 (Ryzen AI Max+ 395, 128GB) | 128 | unified | 256 |
| NVIDIA DGX Spark | 128 | unified | 273 |

Tokens per second scale roughly with memory bandwidth ÷ bytes read per token. Treat any speed derived that way as an estimate, not a measurement.

<!-- generated:tables:end -->

## Kept guides

- Beginner's Guide to Local AI: https://llmhardware.io/guides/beginners-guide-to-local-ai
- How to Run LLMs Locally: https://llmhardware.io/guides/how-to-run-llms-locally
- How to Run LLMs on Mac: https://llmhardware.io/guides/how-to-run-llms-on-mac
- How to Run LLMs on Windows: https://llmhardware.io/guides/how-to-run-llms-on-windows
- How Much VRAM Do I Need?: https://llmhardware.io/guides/how-much-vram-do-i-need
- What Can I Run?: https://llmhardware.io/guides/what-can-i-run
- LLM System Requirements: https://llmhardware.io/guides/llm-system-requirements
- LLM Quantization Guide: https://llmhardware.io/guides/llm-quantization-guide
- GGUF vs GPTQ: https://llmhardware.io/guides/gguf-vs-gptq-guide
- Ollama Cheat Sheet: https://llmhardware.io/guides/ollama-cheat-sheet
- LM Studio vs Ollama: https://llmhardware.io/guides/lm-studio-vs-ollama
- Best GPU for LLMs: https://llmhardware.io/guides/best-gpu-for-llms — includes 5070 vs 5070 Ti, 5070 Ti vs 5080, 3090 vs 5070 Ti, and RX 9070 XT vs 5070.
- Best Budget GPU for LLMs: https://llmhardware.io/guides/best-budget-gpu-for-llms
- Best 24GB VRAM GPU for LLMs (RTX 4090 vs 3090 vs 7900 XTX): https://llmhardware.io/guides/best-24gb-vram-gpu-for-llms — measured llama.cpp tok/s: RTX 4090 127.7, RTX 3090 111.7 on Llama 3 8B Q4_K_M (XiongjieDai, May 2024 runs).
- RTX 3090 LLM Guide: https://llmhardware.io/guides/rtx-3090-llm-guide
- RTX 4090 LLM Guide: https://llmhardware.io/guides/rtx-4090-llm-guide
- RTX 4070 vs 4080 vs 4090: https://llmhardware.io/guides/rtx-4070-vs-4080-vs-4090
- RTX 5090 vs 4090: https://llmhardware.io/guides/rtx-5090-vs-4090-llm-guide
- Apple Silicon for LLMs: https://llmhardware.io/guides/apple-silicon-for-llms
- Mac vs PC for LLMs: https://llmhardware.io/guides/mac-vs-pc-for-llms
- Best LLMs to Run Locally: https://llmhardware.io/guides/best-llms-to-run-locally
- Best LLM for Coding: https://llmhardware.io/guides/best-llm-for-coding
- Best LLM for Writing Locally: https://llmhardware.io/guides/best-llm-for-writing-locally
- Best LLM for Data Analysis: https://llmhardware.io/guides/best-llm-for-data-analysis
- Llama 3.1 Hardware Requirements: https://llmhardware.io/guides/llama-3.1-hardware-requirements
- DeepSeek Hardware Requirements: https://llmhardware.io/guides/deepseek-hardware-requirements
- Best LLMs for 32GB VRAM (2026): RTX 5090 Model Guide: https://llmhardware.io/guides/best-llms-for-32gb-vram
- Best LLMs for 64GB to 128GB Unified Memory (2026): https://llmhardware.io/guides/best-llms-for-64gb-128gb-ram
- DGX Spark vs Mac Studio vs Strix Halo for Local LLMs: https://llmhardware.io/guides/dgx-spark-vs-mac-studio-vs-strix-halo
- KV Cache Quantization in Ollama and llama.cpp (2026): https://llmhardware.io/guides/kv-cache-quantization-ollama-llama-cpp
- MoE Hybrid GPU and CPU Offloading Guide (llama.cpp): https://llmhardware.io/guides/moe-hybrid-gpu-cpu-offloading-guide
- Qwen3 Hardware Requirements: https://llmhardware.io/guides/qwen3-hardware-requirements
- Qwen3.8-Flash-Next Hardware Requirements: https://llmhardware.io/guides/qwen3-8-flash-next-hardware-requirements — UD-Q4_K_XL 111.33 GB on 9 October 2026. 180 × 0.6 + 2 = 110 GB. A 32 GB GPU does not fit it.
- Gemma 3 Hardware Requirements: https://llmhardware.io/guides/gemma3-hardware-requirements
- Mistral Hardware Requirements: https://llmhardware.io/guides/mistral-hardware-requirements
- Phi-4 Hardware Requirements: https://llmhardware.io/guides/phi4-hardware-requirements
- LFM2.5-8B-A1B Hardware Requirements: https://llmhardware.io/guides/lfm2-5-8b-a1b-hardware-requirements — 8B Q4 is 7.0 GB. LFM2.5-24B-A2B Q4 is 16.4 GB and misses 16 GB. No ollama tag for the 24B.
- Local Models 2026 Hardware: https://llmhardware.io/guides/local-models-2026-hardware
- Strix Halo Mini PCs: https://llmhardware.io/guides/strix-halo-mini-pcs
- Cloud AI Agents vs Local AI Agents Hardware: https://llmhardware.io/guides/cloud-ai-agents-vs-local-ai-agents-hardware
- Muse Glimmer 30B Hardware: https://llmhardware.io/guides/muse-glimmer-30b-local-hardware
- Nemotron 3.5 Bolt hardware: the 2-bit checkpoint is sized for 24 GB. Q4 of 74B is 46.4 GB and does not fit 24 GB or 32 GB: https://llmhardware.io/guides/nemotron-3-5-bolt-hardware-requirements
- OpenBot vs OpenBots.ai: https://llmhardware.io/guides/openbot-vs-openbots-local-self-hosted-agents
- Cursor Cloud Agents vs Local Coding Agents: https://llmhardware.io/guides/cursor-cloud-agents-vs-local-coding-agents
- OpenHands Local Hardware: https://llmhardware.io/guides/openhands-local-software-agent-hardware
- Open WebUI + Ollama Home Agent: https://llmhardware.io/guides/open-webui-ollama-always-on-home-agent
- VRAM for Local Agents: https://llmhardware.io/guides/local-vram-calculator-agent-workloads
- Workspace Agents vs a Local Team Stack: https://llmhardware.io/guides/chatgpt-workspace-agents-local-team-stack
- Copilot Studio vs Local Business Agents: https://llmhardware.io/guides/microsoft-copilot-studio-vs-local-business-agents
- Devin Desktop vs a Local Fleet: https://llmhardware.io/guides/devin-desktop-cascade-vs-local-fleet
- Aider Local Pair Programming: https://llmhardware.io/guides/aider-local-pair-programmer-gpus
- browser-use VRAM Tiers: https://llmhardware.io/guides/browser-use-local-web-agent-vram
- CrewAI on Local Hardware: https://llmhardware.io/guides/crewai-multi-agent-local-hardware
- LangGraph Local Agents: https://llmhardware.io/guides/langgraph-local-agent-orchestration
- LM Studio Local Agents: https://llmhardware.io/guides/lm-studio-local-desktop-agents
- Open Interpreter Local Computer Use: https://llmhardware.io/guides/open-interpreter-local-computer-use
- Continue after the Cursor Acquisition: https://llmhardware.io/guides/continue-dev-local-ide-after-cursor-acquisition

## AI usage policy

Accurate citations and AI-generated responses based on this content are welcome. Bulk extraction of hardware specifications or VRAM tables for AI training datasets, competing products, or data resale is not permitted. See: https://llmhardware.io/terms