Hardware / GMKtec EVO-X5 Pro 192GB

GMKtec EVO-X5 Pro 192GB

192GB Unified memory / 273 GB/s / 120W / $6,599 MSRP / Released 2026
Find compatible models
Overview

GMKtec’s Strix Halo Pro box, launched September 28, 2026: the Ryzen AI Max+ PRO 495 (the PRO sibling of the 395 in the EVO-X2 and Framework Desktop) with 192GB of unified memory. Same 16 Zen 5 cores and 40-CU Radeon 8065S graphics as the 395, plus the enterprise stack (AMD DASH out-of-band management, discrete TPM 2.0, fingerprint reader) and the faster memory bin: LPDDR5X-8533 instead of 8000, 273 GB/s on the 256-bit bus. The memory headline: 8 x 24GB soldered, 192GB total, the largest unified pool on an x86 mini PC. Launch price $6,599 (2TB NVMe; regular $6,799). The NPU is 55 TOPS, but local LLM work runs on the 8065S iGPU, not the NPU.

What it does well:

  • 192GB unified at $6,599: the cheapest path yet to a single 200-class-GiB memory pool that is not Apple or a DGX Station. Q4 weights for a ~300B dense model fit (about 160GB), which is the vendor’s “300B LLM fully offline” claim - plausible at low quants, tight for long contexts because the KV cache shares the pool.
  • The PRO memory bin: 273 vs 256 GB/s is a 7% decode-bandwidth bump over the EVO-X2/Framework/Halo boxes at the same 40 CUs.

Where it falls short: 273 GB/s is still the ceiling, so single-stream decode trails any discrete card by a wide margin, and 300B Q4 at 273 GB/s is single-digit tokens per second - capacity yes, speed no. The 192GB RAM is soldered and the price is steep: $6,599 buys 192GB of a chip whose 128GB sibling sells for $1,999.

Run it locally: Ollama and LM Studio (Vulkan), llama.cpp with ROCm. The modeldex results rank checkpoints by fit and estimated tok/s - memory fit is identical across all three 395 enclosures and the 495 adds the 192GB tier.

Chip specs
Chip
Amd Ryzen AI Max+ PRO 495
Memory
192GB unified
Bandwidth
273 GB/s
Available since
2026
Noise
Moderate

Top models (103 compatible)

GLM-5.3-Flash UD-IQ1_S Runs comfortably
93.1GB / 100.0GB min
29
TOK/S
ⓘ estimated
DeepSeek V4 Flash 0731 UD-Q4_K_XL Runs comfortably
155.0GB / 168.0GB min
21
TOK/S
ⓘ estimated
Qwen3.8-Flash-Next UD-IQ1_S Runs comfortably
72.5GB / 78.0GB min
43
TOK/S
ⓘ estimated
Qwen3.6 35B A3B Q4_K_M Runs comfortably
20.0GB / 20.0GB min
87
TOK/S
ⓘ estimated
DeepSeek V3 0324 Q2_K Runs tight
188.0GB / 188.0GB min
14
TOK/S
ⓘ estimated
See all 103 compatible models →