A Reddit tech user released benchmark data this week: DeepSeek's latest V4 Flash model ran at 22-28 tokens/sec on a 128GB home mini PC. That number means local LLMs have finally moved from "theoretically runnable" to "daily-use capable."
What This Is
The tested model is DeepSeek's V4 Flash — the lightweight, speed-focused variant of the V4 family. The platform was AMD's Strix Halo mini PC, heavily promoted last year, notable for unified memory shared between CPU and GPU that removes VRAM bottlenecks during generation.
The tester used an acceleration technique called speculative decoding (think of it as: a small model drafts the next passage, the large model only verifies it), pushing generation speed from ~20 tokens/sec to around 28 tokens/sec. The gain looks modest, but local deployment often lives exactly on this edge between "barely usable" and "unusably laggy."
Industry View
The call worth watching: Chinese open-source LLMs are transitioning from "big parameters" to "actually runs locally."
Bulls argue that top-tier open-source models like DeepSeek, paired with community optimization, make local deployment genuinely meaningful for indie developers, small companies, and even individuals. For data-resident scenarios (medical records, internal code, customer contracts), local is a visible path forward.
But we want to flag an often-overlooked cost ledger: the test mini PC costs roughly 15,000–20,000 RMB, plus Linux setup, command-line tuning, and model file maintenance — not something most enterprise IT teams will willingly pick up. In the short term, "runs locally" and "most people willing to run it locally" remain far apart. Hyperscaler compute revenue may actually be reinforced by this trend: local runners are the exception; the vast majority of enterprises still depend on the cloud.
Impact on Regular People
- Enterprise IT: Highly regulated industries like finance and healthcare gain another "local LLM" option, but high cost and ops overhead keep it from going mainstream anytime soon.
- Individual professionals: Power users willing to tinker and concerned about data privacy can start paying attention; for most, ChatGPT, ERNIE Bot, and Tongyi remain the more realistic choice.
- Consumer market: Large-memory mini PCs are forming a new product category, currently paid for mostly by developers and enthusiasts — far from mass-market adoption.