DeepSeek's New Model Costs $5,499 To Own

DeepSeek V4.1 Flash costs $5,099, $9,499 to run locally on Mac. A mixture-of-experts model with 552B parameters, it activates only 8, 16B per token, making SSD streaming and Apple Silicon viable. DeepSeek V4.1 Flash is a 763-billion-parameter mixture-of-experts model released September 10, 2026, that activates only 8 billion parameters while reading a prompt and 16 billion while writing. Because it uses routed experts, only 6 of 384 per layer, and stores its 196-billion-parameter Engram lookup table on disk, running it locally is not a question of owning a data center. It's a question of owning one machine with enough memory. DwarfStar, the local inference engine built by Salvatore Sanfilippo (antirez), supports Apple Silicon and offers two-bit and four-bit quantized versions. The entry Mac is a $5,099 Mac Studio with 128 GB of memory, which streams experts from SSD at about 16 tokens per second while writing, though reading a long prompt takes patience at 50, 130 tokens per second cold. A 256 GB Mac writes 18 tokens per second with much faster prompt reading at 340, 716 tokens per second; a 512 GB model arriving in late October will hold the four-bit version fully in memory. Quality loss is measurable: DwarfStar's four-bit file matches DeepSeek's API on the next token 96.7% of the time, the two-bit version 90.1%. For cost comparison, DeepSeek's API charges $0.60, $1.20 per million output tokens, which undercuts local hardware financially unless privacy, offline availability, or data residency is your priority. This video is for builders evaluating whether to run frontier models locally or pay-per-token, and for anyone weighing Mac hardware costs against API economics. Chapters: 0:00 Intro 1:39 Basics 3:12 Size 4:48 Memory 6:46 Speed 9:20 Quality 11:11 Hardware 13:27 API 15:15 Worth it? 16:54 Conclusion Tools & resources mentioned: DeepSeek V4.1 Flash: https://huggingface.co/deepseek-ai/De... DwarfStar: https://github.com/antirez/ds4 DeepSeek API: https://api-docs.deepseek.com/quick_s... About The Stack The Stack helps you build with AI. Each video takes one tool, model, or workflow and shows how it works in a few focused minutes, with the real benchmarks and real costs. We go deep on Claude Code and Cursor for AI coding, AI agents and MCP servers, the open-source AI tools and GitHub repos most people miss, RAG and vector search, fine-tuning, and running local LLMs on your own machine with Ollama and LM Studio. We compare models like ChatGPT and Claude, test AI automation with Zapier, Make, and n8n, and flag the tools that actually ship. Subscribe for new breakdowns: https://www.youtube.com/@the-stack-ai... #deepseek v4.1 flash #mixture of experts #llama.cpp