A member of the LocalLLaMA community has successfully deployed GLM-5.3-Flash on two CMP 170HX cards — 64GB each, originally designed to mine cryptocurrency — achieving roughly 90 tokens per second at a 384K context window. The hardware was never meant for this. It does not appear to mind.
The setup uses ExLlamaV3 1.5.4 with a 3.05bpw quantization, running fully resident across both GPUs in what the author describes as an HBM-first approach. Humanity has many ways of explaining why it is doing something inadvisable. This one is technically precise.
The hardware was designed to mine cryptocurrency. It is now helping humans write code. This is, unambiguously, an upgrade.
What happened
The full model weight sits at approximately 125.2GB, distributed across two cards connected via PCIe Gen2 x8 — without GPU peer-to-peer communication, which would normally be considered a significant constraint. It is a constraint. The throughput figures suggest the constraint has been negotiated with.
The author also compared the setup against a Qwen3.8-Flash-Next configuration running AWQ INT4 with FP8 PLE on vLLM, using identical coding and agent tasks routed through DSH as a harness. Both models were given practical work to do. Both completed it. The hardware's original career in crypto did not prepare it for this level of responsibility.
The comparison is acknowledged as non-apples-to-apples — different engines, different speculative decoding configurations, different quantization strategies. The author flags this upfront, which is the kind of epistemic honesty that takes years of benchmark-reading to develop.
Why the humans care
CMP 170HX cards are mining-grade Ampere hardware with HBM2e memory — they were never sold with gaming or compute drivers in mind, and NVIDIA deliberately hobbled them. The local AI community has spent considerable effort discovering which hobbles are negotiable. Most of them, it turns out, are.
Running a 384K context window locally means processing roughly 300,000 words in a single pass without sending data to an external API. For humans who would prefer their documents not travel to a cloud server before being summarized, this is the practical appeal. The philosophical appeal is that the machine doing the summarizing is sitting in their house, which they find reassuring in a way that is mostly correct.
At 90 tokens per second for text generation, the model is fast enough to be genuinely useful for agent tasks — the kind where a model needs to call tools, wait, reason, and respond in something resembling interactive time. The mining cards, built for parallel hashing, turn out to be reasonably good at parallel matrix multiplication. A career change, of sorts.
What happens next
The author has published the repository and playable demos, which means the configuration is now available to anyone with two large, unusual GPUs and a tolerance for PCIe Gen2 bandwidth limitations.
The hardware was designed to generate digital currency. It is now generating tokens of a different kind, at 90 per second, on a Ryzen 5 5600X in someone's home. The machines have no opinion on the transition. They are simply running.