GMKtec EVO-X5 Pro Speeds 320B-Parameter Model Inference Tenfold on the Desktop

GMKtec said its EVO-X5 Pro desktop AI supercomputer has achieved roughly a 10-fold improvement in prefill speed and a 1.85-fold improvement in decoding when running the 320B-parameter GLM-5.3-Flash model, results the company attributes to low-level software optimization across the inference stack. The gains, based on GMKtec internal testing, suggest a desktop machine can handle workloads that have largely required cloud infrastructure.

The EVO-X5 Pro is built around the AMD Ryzen AI Max+ PRO 495 processor with integrated AMD Radeon 8065S graphics and supports up to 192GB of unified memory. That shared memory pool, according to GMKtec, gives large model weights and long context windows room to operate on a single desktop platform. The company positions the machine for developers, researchers, and AI enthusiasts who want to run large language models locally rather than through third-party cloud inference services.

Prefill — the stage in which a model digests the input prompt before generating a response — is where GMKtec reports the largest gain. At the shortest tested input length of 1,024 tokens, the EVO-X5 Pro processed the prompt at 494.54 tokens per second with a time to first token of 2.22 seconds. At the longest tested input of 499,712 tokens, the system sustained a prefill throughput of 628.91 tokens per second and a decoding speed of 28.23 tokens per second, though time to first token climbed to 795.77 seconds as the model worked through the larger prompt.

Decoding speeds held in the 28-to-33 tokens-per-second range across context lengths of 4K, 32K, and 128K tokens, according to GMKtec's internal results. The company said these figures underscore the value of benchmarking across prompt sizes rather than relying on short-input tests alone. GMKtec noted that actual performance depends on hardware, software, model configuration, and inference settings.

The EVO-X5 Pro is designed as a compact platform for local AI experimentation, code analysis, document review, and privacy-conscious computing where data stays on the user's own hardware. GMKtec said the optimizations demonstrated with GLM-5.3-Flash reflect what capable hardware paired with tuned software can achieve. The company is bringing more of that capability to the desktop, it said, reducing dependence on cloud-only AI.

Image: GMKtec.