# VRAMGlass

> VRAMGlass (vramglass.com) is a free reference for running open-weight LLMs on your own hardware. It tracks asking prices for 57 GPUs and AI workstations on eBay (US), Xianyu (China) and Amazon.co.jp (Japan), and calculates how much VRAM 45 versions of 9 model families need at each quantization and context length.

Coverage spans NVIDIA, AMD, Intel and Apple hardware, including unified-memory machines such as the DGX Spark and Mac Studio and aftermarket memory-modded cards such as the RTX 4090 48GB and RTX 2080 Ti 22GB. Pages are server-rendered HTML in English, Simplified Chinese (/zh) and Japanese (/ja), and need no login. The text of every GPU and model page (specs, buying notes, VRAM tables and FAQs) is collected in [llms-full.txt](https://vramglass.com/llms-full.txt); current prices are only on the GPU pages.

## What VRAMGlass answers

- What a GPU costs right now on each marketplace, and its price history: each GPU page, for example [GeForce RTX 3090](https://vramglass.com/gpu/rtx-3090).
- How much VRAM a model needs at 2-bit, 4-bit and 8-bit and at longer contexts, and which single-GPU, multi-GPU or MoE-offload setups run it: each model version page, for example [Qwen3.8 27B](https://vramglass.com/model/qwen/qwen3-8-27b).
- Which models a GPU can run and roughly how fast: each GPU page.
- Whether GPU prices overall are rising or falling: the [GPU price index](https://vramglass.com/index).
- What a DDR4 or DDR5 memory stick costs (desktop UDIMM and server RDIMM, per stick and per GB, with 7-day and 30-day changes), and which server builds run large MoE models from system RAM: [DDR4 prices](https://vramglass.com/ddr4) and [DDR5 prices](https://vramglass.com/ddr5).

## Citing VRAMGlass

- Cite the page URL without query parameters. English pages show eBay (US) prices in USD, /zh pages show Xianyu prices in CNY and /ja pages show Amazon.co.jp prices in JPY. Adding `?market=ebay`, `?market=xianyu` or `?market=amazon_jp` only switches the market shown on the same page; it is not a separate page.
- Prices are medians of current asking prices, not sold prices. A price needs at least 3 valid listings, collection runs about every 6 hours, and every price carries its collection time. Quote a price together with its date and marketplace.
- VRAM needs are weights + KV cache + runtime overhead, in decimal GB to match Hugging Face file sizes. Token speeds are estimated from memory bandwidth, not benchmarks. Formulas and limits: [Methodology](https://vramglass.com/methodology).

## MCP server for AI agents

- VRAMGlass runs a free remote MCP server at https://mcp.vramglass.com/mcp (Streamable HTTP) with the same data as the site: current GPU and DDR4/DDR5 memory prices, price history, the price index, VRAM needs, the PC builds shown on model pages, the same rule-filtered eBay deals shown on GPU and memory pages, and GPU rental prices.
- It needs a free VRAMGlass account. Tools that can send a header use an API key (`Authorization: Bearer <key>`, created under Your account → MCP connections); Claude and ChatGPT connectors sign in with OAuth instead.
- Setup steps and the tool list: [Developer docs](https://vramglass.com/docs#mcp). Overview: [VRAMGlass MCP](https://vramglass.com/mcp).

## GPUs and AI workstations

### NVIDIA

- [DGX Spark 128GB](https://vramglass.com/gpu/dgx-spark): 128 GB unified memory (LPDDR5X), 273 GB/s
- [RTX PRO 6000 Blackwell](https://vramglass.com/gpu/rtx-pro-6000-blackwell): 96 GB VRAM (GDDR7 ECC), 1792 GB/s
- [RTX PRO 5500 Blackwell](https://vramglass.com/gpu/rtx-pro-5500-blackwell): 84 GB VRAM (GDDR7 ECC), 1398 GB/s
- [RTX PRO 6000D](https://vramglass.com/gpu/rtx-pro-6000d): 84 GB VRAM (GDDR7 ECC), 1398 GB/s
- [A100 80GB PCIe](https://vramglass.com/gpu/a100-80gb-pcie): 80 GB VRAM (HBM2e), 1935 GB/s
- [H100 PCIe](https://vramglass.com/gpu/h100-pcie): 80 GB VRAM (HBM2e), 2000 GB/s
- [RTX PRO 5000 Blackwell 72GB](https://vramglass.com/gpu/rtx-pro-5000-blackwell-72g): 72 GB VRAM (GDDR7 ECC), 1344 GB/s
- [CMP 170HX 64GB](https://vramglass.com/gpu/cmp-170hx-64g-mod): 64 GB VRAM (HBM2e), 1493 GB/s, modified card (not an official configuration)
- [GeForce RTX 4090 48GB](https://vramglass.com/gpu/rtx-4090-48g-mod): 48 GB VRAM (GDDR6X), 1008 GB/s, modified card (not an official configuration)
- [GeForce RTX 4090D 48GB](https://vramglass.com/gpu/rtx-4090d-48g-mod): 48 GB VRAM (GDDR6X), 1008 GB/s, modified card (not an official configuration)
- [L40S](https://vramglass.com/gpu/l40s): 48 GB VRAM (GDDR6), 864 GB/s
- [RTX 6000 Ada](https://vramglass.com/gpu/rtx-6000-ada): 48 GB VRAM (GDDR6 ECC), 960 GB/s
- [RTX A6000](https://vramglass.com/gpu/rtx-a6000): 48 GB VRAM (GDDR6 ECC), 768 GB/s
- [RTX PRO 5000 Blackwell 48GB](https://vramglass.com/gpu/rtx-pro-5000-blackwell-48g): 48 GB VRAM (GDDR7 ECC), 1344 GB/s
- [A100 40GB PCIe](https://vramglass.com/gpu/a100-40gb-pcie): 40 GB VRAM (HBM2), 1555 GB/s
- [CMP 170HX 40GB](https://vramglass.com/gpu/cmp-170hx-40g-mod): 40 GB VRAM (HBM2e), 1560 GB/s, modified card (not an official configuration)
- [GeForce RTX 4080 SUPER 32GB](https://vramglass.com/gpu/rtx-4080-super-32g-mod): 32 GB VRAM (GDDR6X), 736.3 GB/s, modified card (not an official configuration)
- [GeForce RTX 5090](https://vramglass.com/gpu/rtx-5090): 32 GB VRAM (GDDR7), 1792 GB/s
- [GeForce RTX 5090 D](https://vramglass.com/gpu/rtx-5090-d-v1): 32 GB VRAM (GDDR7), 1792 GB/s
- [RTX PRO 4500 Blackwell](https://vramglass.com/gpu/rtx-pro-4500-blackwell): 32 GB VRAM (GDDR7 ECC), 896 GB/s
- [Tesla V100 32GB](https://vramglass.com/gpu/tesla-v100-32gb): 32 GB VRAM (HBM2), 900 GB/s
- [GeForce RTX 3090](https://vramglass.com/gpu/rtx-3090): 24 GB VRAM (GDDR6X), 936.2 GB/s
- [GeForce RTX 4090](https://vramglass.com/gpu/rtx-4090): 24 GB VRAM (GDDR6X), 1008 GB/s
- [GeForce RTX 5090 D V2](https://vramglass.com/gpu/rtx-5090-d-v2): 24 GB VRAM (GDDR7), 1344 GB/s
- [RTX PRO 4000 Blackwell](https://vramglass.com/gpu/rtx-pro-4000-blackwell): 24 GB VRAM (GDDR7 ECC), 672 GB/s
- [Tesla P40](https://vramglass.com/gpu/tesla-p40): 24 GB VRAM (GDDR5), 347 GB/s
- [GeForce RTX 2080 Ti 22GB](https://vramglass.com/gpu/rtx-2080-ti-22g-mod): 22 GB VRAM (GDDR6), 616 GB/s, modified card (not an official configuration)
- [GeForce RTX 4060 Ti 16GB](https://vramglass.com/gpu/rtx-4060-ti-16gb): 16 GB VRAM (GDDR6), 288 GB/s
- [GeForce RTX 4080](https://vramglass.com/gpu/rtx-4080): 16 GB VRAM (GDDR6X), 716.8 GB/s
- [GeForce RTX 5060 Ti 16GB](https://vramglass.com/gpu/rtx-5060-ti-16gb): 16 GB VRAM (GDDR7), 448 GB/s
- [GeForce RTX 5070 Ti](https://vramglass.com/gpu/rtx-5070-ti): 16 GB VRAM (GDDR7), 896 GB/s
- [GeForce RTX 5080](https://vramglass.com/gpu/rtx-5080): 16 GB VRAM (GDDR7), 960 GB/s
- [Tesla T10](https://vramglass.com/gpu/tesla-t10): 16 GB VRAM (GDDR6), 403.2 GB/s
- [Tesla V100 16GB](https://vramglass.com/gpu/tesla-v100-16gb): 16 GB VRAM (HBM2), 900 GB/s
- [GeForce RTX 3060 12GB](https://vramglass.com/gpu/rtx-3060-12gb): 12 GB VRAM (GDDR6), 360 GB/s
- [GeForce RTX 5070](https://vramglass.com/gpu/rtx-5070): 12 GB VRAM (GDDR7), 672 GB/s

### AMD

- [Ryzen AI Max+ 395 128GB](https://vramglass.com/gpu/ryzen-ai-max-395-128gb): 128 GB unified memory (LPDDR5X), 256 GB/s
- [Instinct MI210](https://vramglass.com/gpu/instinct-mi210): 64 GB VRAM (HBM2e), 1638.4 GB/s
- [Radeon PRO W7900](https://vramglass.com/gpu/radeon-pro-w7900): 48 GB VRAM (GDDR6 ECC), 864 GB/s
- [Instinct MI100](https://vramglass.com/gpu/instinct-mi100): 32 GB VRAM (HBM2), 1228.8 GB/s
- [Instinct MI50 32GB](https://vramglass.com/gpu/instinct-mi50-32gb): 32 GB VRAM (HBM2), 1024 GB/s
- [Radeon AI PRO R9700](https://vramglass.com/gpu/radeon-ai-pro-r9700): 32 GB VRAM (GDDR6), 640 GB/s
- [Radeon PRO W7800](https://vramglass.com/gpu/radeon-pro-w7800): 32 GB VRAM (GDDR6 ECC), 576 GB/s
- [Radeon RX 7900 XTX](https://vramglass.com/gpu/rx-7900-xtx): 24 GB VRAM (GDDR6), 960 GB/s
- [Radeon RX 7900 XT](https://vramglass.com/gpu/rx-7900-xt): 20 GB VRAM (GDDR6), 800 GB/s
- [Radeon RX 6950 XT](https://vramglass.com/gpu/rx-6950-xt): 16 GB VRAM (GDDR6), 576 GB/s
- [Radeon RX 7800 XT](https://vramglass.com/gpu/rx-7800-xt): 16 GB VRAM (GDDR6), 624.1 GB/s
- [Radeon RX 9070 XT](https://vramglass.com/gpu/rx-9070-xt): 16 GB VRAM (GDDR6), 640 GB/s

### Intel

- [Arc Pro B65](https://vramglass.com/gpu/arc-pro-b65): 32 GB VRAM (GDDR6), 608 GB/s
- [Arc Pro B70](https://vramglass.com/gpu/arc-pro-b70): 32 GB VRAM (GDDR6), 608 GB/s
- [Arc Pro B60](https://vramglass.com/gpu/arc-pro-b60): 24 GB VRAM (GDDR6), 456 GB/s

### Apple

- [Mac Studio M3 Ultra 512GB](https://vramglass.com/gpu/mac-studio-m3-ultra-512gb): 512 GB unified memory (LPDDR5), 819 GB/s
- [Mac Studio M4 Max 128GB](https://vramglass.com/gpu/mac-studio-m4-max-128gb): 128 GB unified memory (LPDDR5X), 546 GB/s
- [Mac Studio M5 Max 128GB](https://vramglass.com/gpu/mac-studio-m5-max-128gb): 128 GB unified memory (LPDDR5X), 614 GB/s
- [MacBook Pro M4 Max 128GB](https://vramglass.com/gpu/macbook-pro-m4-max-128gb): 128 GB unified memory (LPDDR5X), 546 GB/s
- [MacBook Pro M5 Max 128GB](https://vramglass.com/gpu/macbook-pro-m5-max-128gb): 128 GB unified memory (LPDDR5X), 614 GB/s
- [Mac Studio M5 Ultra 96GB](https://vramglass.com/gpu/mac-studio-m5-ultra-96gb): 96 GB unified memory (LPDDR5X), 1228.8 GB/s

## Open-weight models

### DeepSeek (DeepSeek)

- [DeepSeek overview](https://vramglass.com/model/deepseek): every version, licenses and a VRAM table by context length
- [DeepSeek V4.1 Flash](https://vramglass.com/model/deepseek/v4-1-flash): 763B MoE, 16B active; 4-bit needs about 446 GB at 8K context
- [DeepSeek V4 Pro](https://vramglass.com/model/deepseek/v4-pro): 1600B MoE, 49B active; 4-bit needs about 851 GB at 8K context, or 16 GB VRAM plus 837 GB system RAM with MoE expert offloading
- [DeepSeek V4 Flash](https://vramglass.com/model/deepseek/v4-flash): 284B MoE, 13B active; 4-bit needs about 157 GB at 8K context, or 6 GB VRAM plus 152 GB system RAM with MoE expert offloading
- [DeepSeek V3.2](https://vramglass.com/model/deepseek/v3-2): 671B MoE, 37B active; 4-bit needs about 408 GB at 8K context, or 13 GB VRAM plus 396 GB system RAM with MoE expert offloading
- [DeepSeek R1 Qwen3 8B](https://vramglass.com/model/deepseek/r1-qwen3-8b): 8.2B dense; 4-bit needs about 8 GB at 8K context
- [DeepSeek R1 Distill Qwen 14B](https://vramglass.com/model/deepseek/r1-distill-qwen-14b): 14.8B dense; 4-bit needs about 12 GB at 8K context
- [DeepSeek R1 Distill Qwen 32B](https://vramglass.com/model/deepseek/r1-distill-qwen-32b): 32.8B dense; 4-bit needs about 24 GB at 8K context

### Gemma (Google DeepMind)

- [Gemma overview](https://vramglass.com/model/gemma): every version, licenses and a VRAM table by context length
- [Gemma 4 12B](https://vramglass.com/model/gemma/gemma-4-12b): 11.9B dense; 4-bit needs about 9 GB at 8K context
- [Gemma 4 26B A4B](https://vramglass.com/model/gemma/gemma-4-26b-a4b): 25.2B MoE, 3.8B active; 4-bit needs about 19 GB at 8K context, or 5 GB VRAM plus 16 GB system RAM with MoE expert offloading
- [Gemma 4 31B](https://vramglass.com/model/gemma/gemma-4-31b): 30.7B dense; 4-bit needs about 21 GB at 8K context
- [Gemma 4 E4B](https://vramglass.com/model/gemma/gemma-4-e4b): 8B dense; 4-bit needs about 7 GB at 8K context

### GLM (Z.ai / Zhipu AI)

- [GLM overview](https://vramglass.com/model/glm): every version, licenses and a VRAM table by context length
- [GLM 5.3](https://vramglass.com/model/glm/glm-5-3): 744B MoE, 40B active; 4-bit needs about 470 GB at 8K context, or 14 GB VRAM plus 457 GB system RAM with MoE expert offloading
- [GLM 5.3 Flash](https://vramglass.com/model/glm/glm-5-3-flash): 320B MoE, 18B active; 4-bit needs about 201 GB at 8K context, or 9 GB VRAM plus 194 GB system RAM with MoE expert offloading
- [GLM 5.2](https://vramglass.com/model/glm/glm-5-2): 744B MoE, 40B active; 4-bit needs about 468 GB at 8K context, or 14 GB VRAM plus 455 GB system RAM with MoE expert offloading
- [GLM 4.7 Flash](https://vramglass.com/model/glm/glm-4-7-flash): 30B MoE, 3B active; 4-bit needs about 20 GB at 8K context, or 4 GB VRAM plus 18 GB system RAM with MoE expert offloading
- [GLM 4.5 Air](https://vramglass.com/model/glm/glm-4-5-air): 106B MoE, 12B active; 4-bit needs about 76 GB at 8K context, or 8 GB VRAM plus 70 GB system RAM with MoE expert offloading

### gpt-oss (OpenAI)

- [gpt-oss overview](https://vramglass.com/model/gpt-oss): every version, licenses and a VRAM table by context length
- [gpt-oss Safeguard 20B](https://vramglass.com/model/gpt-oss/gpt-oss-safeguard-20b): 20.9B MoE, 3.6B active; 4-bit needs about 13 GB at 8K context, or 3 GB VRAM plus 11 GB system RAM with MoE expert offloading
- [gpt-oss 120B](https://vramglass.com/model/gpt-oss/gpt-oss-120b): 116.8B MoE, 5.1B active; 4-bit needs about 65 GB at 8K context, or 4 GB VRAM plus 62 GB system RAM with MoE expert offloading
- [gpt-oss 20B](https://vramglass.com/model/gpt-oss/gpt-oss-20b): 20.9B MoE, 3.6B active; 4-bit needs about 13 GB at 8K context, or 3 GB VRAM plus 11 GB system RAM with MoE expert offloading

### Hunyuan (Tencent)

- [Hunyuan overview](https://vramglass.com/model/hunyuan): every version, licenses and a VRAM table by context length
- [Hy4 preview](https://vramglass.com/model/hunyuan/hy4-preview): 770B MoE, 49B active; 4-bit needs about 470 GB at 8K context, or 19 GB VRAM plus 452 GB system RAM with MoE expert offloading
- [Hy3](https://vramglass.com/model/hunyuan/hy3): 295B MoE, 21B active; 4-bit needs about 186 GB at 8K context, or 11 GB VRAM plus 177 GB system RAM with MoE expert offloading

### Kimi (Moonshot AI)

- [Kimi overview](https://vramglass.com/model/kimi): every version, licenses and a VRAM table by context length
- [Kimi K3](https://vramglass.com/model/kimi/k3): 2800B MoE, 104B active; 4-bit needs about 1511 GB at 8K context, or 32 GB VRAM plus 1480 GB system RAM with MoE expert offloading
- [Kimi K2.7 Code](https://vramglass.com/model/kimi/k2-7): 1000B MoE, 32B active; 4-bit needs about 586 GB at 8K context, or 10 GB VRAM plus 578 GB system RAM with MoE expert offloading
- [Kimi K2.6](https://vramglass.com/model/kimi/k2-6): 1000B MoE, 32B active; 4-bit needs about 586 GB at 8K context, or 10 GB VRAM plus 578 GB system RAM with MoE expert offloading
- [Kimi K2.5](https://vramglass.com/model/kimi/k2-5): 1000B MoE, 32B active; 4-bit needs about 623 GB at 8K context, or 10 GB VRAM plus 615 GB system RAM with MoE expert offloading
- [Kimi Linear 48B A3B](https://vramglass.com/model/kimi/linear-48b-a3b): 49.1B MoE, 3B active; 4-bit needs about 32 GB at 8K context, or 4 GB VRAM plus 30 GB system RAM with MoE expert offloading
- [Kimi K2](https://vramglass.com/model/kimi/k2): 1000B MoE, 32B active; 4-bit needs about 623 GB at 8K context, or 10 GB VRAM plus 615 GB system RAM with MoE expert offloading

### Meta (Meta AI)

- [Meta overview](https://vramglass.com/model/llama): every version, licenses and a VRAM table by context length
- [Muse Spark Open](https://vramglass.com/model/llama/muse-spark-open-816b-a42b): 816B MoE, 42B active; 4-bit needs about 508 GB at 8K context, or 14 GB VRAM plus 495 GB system RAM with MoE expert offloading
- [Llama 4 Scout 17B 16E](https://vramglass.com/model/llama/llama-4-scout-17b-16e): 108.6B MoE, 17B active; 4-bit needs about 69 GB at 8K context, or 11 GB VRAM plus 59 GB system RAM with MoE expert offloading
- [Llama 3.3 70B](https://vramglass.com/model/llama/llama-3-3-70b): 70.5B dense; 4-bit needs about 47 GB at 8K context
- [Llama 3.2 3B](https://vramglass.com/model/llama/llama-3-2-3b): 3.2B dense; 4-bit needs about 5 GB at 8K context
- [Llama 3.1 8B](https://vramglass.com/model/llama/llama-3-1-8b): 8B dense; 4-bit needs about 8 GB at 8K context

### Mistral (Mistral AI)

- [Mistral overview](https://vramglass.com/model/mistral): every version, licenses and a VRAM table by context length
- [Ministral 3 14B](https://vramglass.com/model/mistral/ministral-3-14b): 13.5B dense; 4-bit needs about 11 GB at 8K context
- [Ministral 3 8B](https://vramglass.com/model/mistral/ministral-3-8b): 8.4B dense; 4-bit needs about 8 GB at 8K context
- [Mistral Small 3.2 24B](https://vramglass.com/model/mistral/mistral-small-3-2-24b): 24B dense; 4-bit needs about 17 GB at 8K context
- [Mistral Nemo](https://vramglass.com/model/mistral/mistral-nemo): 12.3B dense; 4-bit needs about 10 GB at 8K context
- [Mistral 7B v0.3](https://vramglass.com/model/mistral/mistral-7b-v0-3): 7.3B dense; 4-bit needs about 7 GB at 8K context

### Qwen (Alibaba)

- [Qwen overview](https://vramglass.com/model/qwen): every version, licenses and a VRAM table by context length
- [Qwen3.8 Flash Next](https://vramglass.com/model/qwen/qwen3-8-flash-next): 176B MoE, 6B active; 4-bit needs about 113 GB at 8K context, or 5 GB VRAM plus 110 GB system RAM with MoE expert offloading
- [Qwen3.8 27B](https://vramglass.com/model/qwen/qwen3-8-27b): 27.8B dense; 4-bit needs about 19 GB at 8K context
- [Qwen3.8 2.4T A95B](https://vramglass.com/model/qwen/qwen3-8-2-4t-a95b): 2446B MoE, 95B active; 4-bit needs about 1313 GB at 8K context, or 29 GB VRAM plus 1286 GB system RAM with MoE expert offloading
- [Qwen3.6 27B](https://vramglass.com/model/qwen/qwen3-6-27b): 27.8B dense; 4-bit needs about 19 GB at 8K context
- [Qwen3.6 35B A3B](https://vramglass.com/model/qwen/qwen3-6-35b-a3b): 36B MoE, 3B active; 4-bit needs about 24 GB at 8K context, or 4 GB VRAM plus 21 GB system RAM with MoE expert offloading
- [Qwen3.5 9B](https://vramglass.com/model/qwen/qwen3-5-9b): 9.7B dense; 4-bit needs about 8 GB at 8K context
- [Qwen3 32B](https://vramglass.com/model/qwen/qwen3-32b): 32.8B dense; 4-bit needs about 24 GB at 8K context
- [Qwen3 8B](https://vramglass.com/model/qwen/qwen3-8b): 8.2B dense; 4-bit needs about 8 GB at 8K context

## Optional

- [Methodology](https://vramglass.com/methodology): price sources, sample rules, VRAM and speed formulas, known limits
- [About and corrections](https://vramglass.com/about): publisher, contact and corrections policy
- [GPU price widget](https://vramglass.com/widget): embed live GPU price cards on another site
- [Developer docs](https://vramglass.com/docs): MCP setup, API keys, the widget embed code and how to cite the data
- [Simplified Chinese edition](https://vramglass.com/zh): the same data with Xianyu (CNY) prices by default
- [Japanese edition](https://vramglass.com/ja): the same data with Amazon.co.jp (JPY) prices by default
- [Sitemap](https://vramglass.com/sitemap.xml): every canonical page with language alternates