Locally Explains

The VRAM You Need For Every Local AI Model (October 2026)

Every figure in the video, where it came from. Prices were checked on 16 October 2026 and move weekly; speeds are other people’s published measurements, never our own.

Model sizes and memory maths

• Qwen3.5 9B card and config: https://huggingface.co/Qwen/Qwen3.5-9B
• Qwen3.5 9B GGUF sizes: https://huggingface.co/unsloth/Qwen3.5-9B-GGUF
• Qwen3 8B config: https://huggingface.co/Qwen/Qwen3-8B
• Qwen3 14B card and config: https://huggingface.co/Qwen/Qwen3-14B
• Qwen3 14B GGUF sizes: https://huggingface.co/unsloth/Qwen3-14B-GGUF
• Qwen 3.8 27B card and config: https://huggingface.co/Qwen/Qwen3.8-27B
• Qwen 3.8 27B Q4_K_M files: https://huggingface.co/lmstudio-community/Qwen3.8-27B-GGUF, https://huggingface.co/ggml-org/Qwen3.8-27B-GGUF and https://huggingface.co/unsloth/Qwen3.8-27B-GGUF
• Qwen 3.8 27B KV cache maths: https://www.techplained.com/run-qwen-3-8-27b-locally
• Ternary Bonsai 2 27B (5.95 GB, 98.2%): https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf
• Bonsai 2 needs Prism’s llama.cpp fork; Ollama cannot load it: https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf/blob/main/KNOWN_ISSUES.md
• Llama 3.3 70B GGUF sizes: https://huggingface.co/bartowski/Llama-3.3-70B-Instruct-GGUF
• Llama 3.3 70B config: https://huggingface.co/unsloth/Llama-3.3-70B-Instruct
• Llama 70B KV cache in a llama.cpp log (160 MiB at 512 tokens): https://huggingface.co/legraphista/Meta-Llama-3.1-70B-Instruct-IMat-GGUF/blob/252373a45c3345bde22d5be01b45bff696b0ad9f/imatrix.log
• Qwen 3.8 Flash Next card: https://huggingface.co/Qwen/Qwen3.8-Flash-Next
• Qwen 3.8 Flash Next GGUF sizes: https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF
• gpt-oss-120b GGUF sizes: https://huggingface.co/unsloth/gpt-oss-120b-GGUF
• GLM-5.3-Flash card and config: https://huggingface.co/zai-org/GLM-5.3-Flash
• GLM-5.3-Flash GGUF sizes: https://huggingface.co/unsloth/GLM-5.3-Flash-GGUF
• Qwen3.5 397B GGUF sizes: https://huggingface.co/unsloth/Qwen3.5-397B-A17B-GGUF
• Downloads per model (last 30 days): https://huggingface.co/models?sort=downloads

Speeds

• RTX 3060 12GB (Qwen3 8B 55.2, Qwen3 14B 31.2 tok/s): https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3060-12gb/
• Qwen 3.8 27B measured VRAM and RTX 3090 speed: https://www.hardware-corner.net/qwen3-8-27b-hardware-tests/
• Strix Halo, Llama 70B: https://github.com/visorcraft/strix-halo-llm-perf
• Strix Halo, Flash Next: https://llamaperf.com/gpu/strix-halo-128gb and https://github.com/Niko1221/Strata/releases/tag/v0.1.40
• Flash Next on Strix Halo (111 GB file crashes at 32K): https://sleepingrobots.com/dreams/qwen38-flash-next-strix-halo/
• Strata (RTX 5070 12GB + 64GB, SSD streaming): https://github.com/Niko1221/Strata
• RTX 5090 in VRAM vs spilled to RAM, Qwen 3.8 27B: https://www.macstories.net/stories/m5-ultra-mac-studio-review-the-dream-mac-for-local-ai-agents/
• Qwen3 235B on M3 Ultra: https://omlx.ai/benchmarks/dn0dqucy
• Usable speed by task: https://openclawdc.com/blog/usable-tokens-per-second-by-task/

Prices (16 October 2026)

• RTX 5060 8GB, $469.99 at Newegg: https://www.newegg.com/zotac-rtx-5060-amp-geforce-rtx-5060-8gb-graphics-card-double-fans/p/N82E16814500658
• Used RTX 3060 12GB, about $370: https://rigprice.com/gpu/rtx-3060-12gb/
• RX 9060 XT 16GB, $499.99 at Newegg: https://www.newegg.com/powercolor-reaper-rx9060xt-16g-a-radeon-rx-9060-xt-16gb-video-card-double-fans/p/N82E16814131880
• Intel Arc Pro B60 24GB, $649.99 at Newegg: https://www.newegg.com/asrock-intel-arc-pro-b60-24gb-graphics/p/N82E16814930148
• Used RTX 3090, about $1,740: https://rigprice.com/gpu/rtx-3090/
• RTX 5070 12GB, $859.99 at Newegg: https://www.newegg.com/gigabyte-gv-n5070eagleoc-ice-12gd-geforce-rtx-5070-12gb-graphics-card-triple-fans/p/N82E16814932775
• 64GB DDR5-5200 kit, $919.99 at Newegg: https://www.newegg.com/team-group-t-force-delta-rgb-64gb-ddr5-5200-cas-latency-cl40-desktop-memory-black/p/N82E16820331982
• GMKtec EVO-X2 64GB $2,199.99 and 128GB $3,499.99: https://www.gmktec.com/products/amd-ryzen%E2%84%A2-ai-max-395-evo-x2-ai-mini-pc
• Bosgame M5 128GB, $3,099.76 pre-order: https://www.bosgamepc.com/products/bosgame-m5-ai-mini-desktop-ryzen-ai-max-395
• Mac Studio M5 Ultra 256GB, $9,499: https://www.apple.com/shop/buy-mac/mac-studio/m5-ultra-chip-30-core-cpu-64-core-gpu-256gb-memory-1tb-storage
• Mac Studio M5 Ultra 512GB option “coming late October”, no price: https://www.apple.com/shop/buy-mac/mac-studio