GMKtec Says Its Halogen Engine Triples Token Speed on EVO-X5 Pro Mini PC

GMKtec says its EVO-X5 Pro mini PC can sustain 75 tokens per second when running Qwen3.8-Flash-Next through a specialized inference engine called Halogen — more than three times the speed the company measured using a general-purpose engine on the same hardware. The result, according to GMKtec, comes entirely from software optimization rather than any change to the underlying AMD Strix Halo platform, which provides 256 GB/s of memory bandwidth.

The EVO-X5 Pro ships with up to 192GB of unified memory and an AMD Ryzen AI Max+ PRO 495 processor. GMKtec positions the machine for users who run large language models locally, including developers working with coding assistants, long-context document analysis, and local AI agents. AMD's Strix Halo platform has drawn criticism for memory bandwidth constraints that limit inference speed, but GMKtec argues that software plays the decisive role in how much of the hardware's potential gets used.

"Hardware sets the limits. Software determines how much of that potential you can actually unlock."

Halogen targets the gfx1151 GPU architecture found in Strix Halo and uses hardware-specific kernels and model-aware execution paths rather than the broad compatibility approach of general-purpose engines such as llama.cpp. The engine is built around the execution patterns of specific model families, optimizing expert routing and data movement for Mixture-of-Experts architectures like Qwen3.8-Flash-Next, which uses close to 180 billion total parameters. MoE models route each input through a selected subset of specialized experts rather than activating every parameter, making memory movement a critical performance factor on bandwidth-constrained platforms.

Halogen also employs a 5.53 bits-per-weight quantization configuration, which GMKtec says balances compact weight storage against model quality without forcing the entire model into a uniform 4-bit representation. The engine uses speculative decoding, in which a smaller draft component proposes candidate tokens that the main model verifies in parallel, allowing multiple tokens to be accepted for the computational cost of fewer sequential steps.

In GMKtec's internal testing, Halogen reached a peak of 94 tokens per second in repetitive-content scenarios such as code completion and structured text generation. The company also reported prompt processing speeds of up to 2,000 tokens per second, enabling the EVO-X5 Pro to ingest large inputs — including contexts of up to 128K tokens — quickly before generating responses.

GMKtec notes that the figures reflect internal testing under specified configurations and that actual performance varies with model settings, context length, quantization, and workload. The EVO-X5 Pro is available now, and GMKtec describes the machine as built for users who want to run demanding AI workloads on desktop hardware without relying on cloud-based inference services.

Image: GMKtec.