This week, a high-upvote post in the LocalLLaMA community boiled down to one data point: a machine assembled for under €1000 (roughly ¥7,800 RMB) can run a 35-billion-parameter LLM locally. We think this deserves attention because it pulls "local AI" out of the geek-toy category and turns it into a budget option an ordinary consumer can actually calculate.

What this is

The core of the setup is using the AMD Ryzen processor's integrated Radeon 780M GPU (iGPU, the graphics chip built into the CPU) as the primary compute, then carving out 48GB of system memory as "VRAM" through kernel parameters, paired with llama.cpp (a community tool for running LLMs on ordinary hardware) and a Vulkan backend (a general-purpose graphics interface that lets the GPU run AI compute) to load open-source models at the scale of Qwen3.5 35B and Gemma4 31B.

Benchmarks posted in the original thread: roughly 21 tokens/s text generation on the 35B model, and around 16 tokens/s on the 30B model using MTP (Multi-Token Prediction, letting the model guess multiple tokens at once to speed up output). That means local conversational response is usable — not "runs but too slow to be functional."

The hardware list is straightforward: a used mini PC at €300–400, two 32GB DDR5 SO-DIMM sticks around €500, an SSD at €50–100 — total roughly ¥7,000–8,000 RMB.

Industry view

The favorable read: once this path works, cloud LLM subscription fees and data privacy concerns have a credible hedge. For SMBs, law firms, and medical institutions with strict cross-border data requirements, locally deploying a 30B-class model is already enough for drafting emails, summarizing documents, and editing contracts.

But the dissent and reservations deserve equal airtime. First, 16–21 tokens/s is an order of magnitude behind what ChatGPT's backend feels like — there's still real distance between "usable" and "pleasant." Second, 35B-scale models still underperform closed-source models like GPT-4 on complex Chinese-language tasks; going local doesn't equal replacing the cloud. Third, this approach depends on geek-tier kernel tuning and Linux know-how that ordinary users can't handle — the supporting ecosystem hasn't grown yet. Fourth, DDR5 memory is itself a price-volatile component; the 48GB-as-VRAM math only holds while memory stays cheap. If prices spike, this calculation falls apart again.

Impact on regular people

For enterprise IT: For industries with strict data-compliance demands, local deployment now has a price tag a CFO can sign off on. But watch out: deployment and maintenance costs go well beyond the hardware line item.

For working professionals: If you're willing to spend a day or two wrestling with Linux and the command line, local AI can now serve as an "offline writing assistant" for anyone handling sensitive documents. Most ordinary office workers still have no reason to self-host — a cloud subscription is the better deal.

For the consumer market: The real signal in this thread is that AI compute is trickling down faster than expected. Hardware vendors and memory makers may fight over the "mini AI workstation" category in 2026 — a product line worth watching.