DeepSeek V4.1 Flash 1017 local: what it actually costs to run..

DeepSeek V4.1 Flash 1017 local, priced on five home machines: what it actually costs in electricity per million tokens, against DeepSeek's own 60 cent API. Should you run DeepSeek locally, or rent it? DeepSeek's new Flash checkpoint is free to download (MIT), and Project Maya now runs it on a single RTX 3090 with 64 GB of RAM. We priced it on five setups people really own: one RTX 3090 + 64 GB (about 10 tokens/s, up to $2.53 per million tokens), the 2014 Dell with two RTX 3060s and 256 GB (9 to 11 tokens/s, up to $2.48), an RTX 5090 + 128 GB (21 tokens/s is the warm-cache number; new content runs 5 to 10, up to $4.07), a Mac Studio with 256 GB (9.46 tokens/s on a 2-bit build, up to $1.44) and two RTX 3090s + 128 GB (about 32 tokens/s, up to $1.34). Plus the 12-watt rule: for every token per second, your PC can draw about 12 watts before DeepSeek's off-peak price is cheaper. Also: the 73-minute cat priced, the box cost, why the API is so cheap (batching), and what 1017 really changed (16% fewer tokens on Artificial Analysis). Watch next: Is local AI worth it? We did the math vs Claude Haiku 5.5..    • Is local AI worth it? We did the math vs C...   Chapters 0:00 DeepSeek V4.1 Flash 1017 local: free model, real power bill 0:32 What 1017 changed, what DeepSeek charges, why it's so big 1:47 The 12-watt rule: when your PC's power bill beats DeepSeek's API 4:00 Machine 1: one RTX 3090 with 64 GB of RAM 6:20 Machine 2: the 2014 Dell with 256 GB of RAM (Project Maya) 8:01 Machine 3: RTX 5090 with 128 GB, and the warm-cache number 9:12 Halfway: why every PC so far loses to 60 cents 10:01 Machine 4: Mac Studio with 256 GB 11:20 Machine 5: two RTX 3090s with 128 GB, the fastest home setup 12:25 The cost nobody prices: the box, and the copy you run 13:34 Why DeepSeek's API is so cheap: batching 15:23 What DeepSeek V4.1 Flash 1017 really changed: 16% fewer tokens 16:32 So, your move: should you run DeepSeek V4.1 Flash 1017 locally? 17:43 The verdict: what DeepSeek V4.1 Flash 1017 costs at home Sources HF: DeepSeek-V4.1-Flash-1017 card: https://huggingface.co/deepseek-ai/De... DeepSeek API: Models & Pricing: https://api-docs.deepseek.com/quick_s... The Next Web: DeepSeek V4.1-Flash: https://thenextweb.com/news/deepseek-... Artificial Analysis: V4.1 Flash 1017: https://artificialanalysis.ai/models/... Artificial Analysis: V4.1 Flash (Sep): https://artificialanalysis.ai/models/... Project Maya README: https://github.com/mw00/project-maya Project Maya v1.1.16: https://github.com/mw00/project-maya/... Project Maya v1.1.17: https://github.com/mw00/project-maya/... Project Maya v1.1.18: https://github.com/mw00/project-maya/... MindStudio: Project Maya review: https://www.mindstudio.ai/blog/glm-5-... RTX 5090 report (JigSawPT): https://github.com/JigSawPT/deepseek-... Mac Studio 512 GB MLX bench: https://github.com/LocalAiCherry/deep... 2-bit MLX on 256 GB Mac: https://huggingface.co/0xSojalSec/Dee... Apple: Mac Studio power: https://support.apple.com/en-us/102027 Notebookcheck: 512 GB Mac Studio retired: https://www.notebookcheck.net/Apple-i... Apple Newsroom: Mac Studio M3 Ultra: https://www.apple.com/newsroom/2025/0... MacRumors: M5 Ultra 512 GB in October: https://www.macrumors.com/2026/08/25/... EIA: Short-Term Energy Outlook: https://www.eia.gov/outlooks/steo/rep... NVIDIA: RTX 3090 specs: https://www.nvidia.com/en-us/geforce/... NVIDIA: RTX 3060 specs: https://www.nvidia.com/en-us/geforce/... NVIDIA: RTX 5090 specs: https://www.nvidia.com/en-us/geforce/... DeepSeek: inference system overview: https://github.com/deepseek-ai/open-i... DeepSeek privacy policy: https://cdn.deepseek.com/policies/en-... Morgans Code: Project Maya video (comments):    • Is $1,500 All You Need to Run DeepSeek V4....   Bruno Vega & Kai: cost to run 1017:    • What Would it Cost to Run Deepseek V4.1 Fl...   Local AI is worth it? (our episode):    • Is local AI worth it? We did the math vs C...   Subscribe:    / @kernelbrief   #DeepSeekV41Flash1017 #LocalLLM #DeepSeek