$35,000 RMB (~$5,000 USD), a single AMD DIY workstation, and the Qwen 122B model hitting 49 tokens/sec—this is this week's concrete scorecard from Reddit's r/LocalLLaMA. What matters isn't how powerful the hardware is, but that running large models locally now has, for the first time, a workstation-grade cost-performance option.
What This Is
The builder paired Framework's Strix Halo motherboard (AMD's single-chip design integrating CPU, GPU, and memory, with unified memory up to 128GB) with an R9700 discrete GPU (AMD's consumer-grade accelerator targeting AI inference). Total build cost: roughly $5,000 USD.
Running Qwen-3.5 122B (Alibaba's open-source MoE large model—"Mixture of Experts" means the model is split into blocks, with only a portion activated per inference), generation speed hit 49 tokens/sec—double the 24 tokens/sec achieved with Strix Halo alone. Prefill speed (reading the entire input before generating begins) at 32K context (roughly 50,000 Chinese characters) improved 2.5x.
The key is the model-splitting approach: the always-needed "dense layers," KV cache (the model's short-term memory of conversation history), and the drafter model (a speedup trick where a small model guesses tokens first, then the large model verifies) live in the R9700's VRAM; the less-frequently-used "expert layers" sit in Strix Halo's large unified memory. Only about 12KB per second needs to flow between the two pieces of hardware—standard PCIe lanes handle it fine.
Industry View
Supporters see a trend signal: when a $35K RMB workstation can run 122B-class models, the cost curve for "local AI" drops from "data-center-grade" to "high-end PC-grade." Data-sensitive industries—law firms, healthcare, manufacturing process data—for the first time have realistic cost-performance economics for building their own on-premise inference machines. NVIDIA's DGX Spark at the same tier is also aimed squarely at this market.
But skepticism runs deep. First, the author himself is a seasoned hardware enthusiast—he describes the tuning process as "took some brainwork"—a typical enterprise IT team can't easily replicate it. Second, the entire setup depends on AMD's ROCm software ecosystem (AMD's in-house AI acceleration framework), which lags significantly behind NVIDIA's CUDA on compatibility; swap in a different model and it might break. Third, this clever split only works for MoE architectures—dense architectures like the Llama series cannot use it.
Impact on Regular People
For enterprise IT: When evaluating "local AI" options this year, put AMD-platform workstations on the comparison list rather than defaulting to NVIDIA-only.
For working professionals: If you're not an AI engineer or hardware enthusiast, this doesn't directly affect you—yet. But it signals that within three to five years, "running a decent AI assistant on your own computer" will shift from a geek toy to a viable option.
For the consumer market: Home AI workstations are moving from PowerPoint slides to real assembly. In 2026 we expect to see more branded complete-machine offerings in the 20,000–40,000 RMB range.