RLX Infra // Measured Public Validation

Different hardware. Same foundation.

RLX Infra has been measured across two substantially different test systems to validate local execution across both accelerated and CPU-only environments. The measured results are public; the internal implementation that produces them remains proprietary.

MEASURED RESULTS // PROPRIETARY METHODS

Measured validation results — not a build guide.

The figures on this page are measured snapshots from two separate RLX Infra validation systems. They are shown to demonstrate observed local performance, not to publish the underlying architecture. The two profiles use different models and workloads, so their numbers should not be read as a direct head-to-head benchmark. Routing logic, allocation strategy, model handling, scheduling, caching and optimization methods remain private to Realystic™ / RLX.

132.7tok/s

Measured decode

GPU validation run

16,987.4tok/s

Measured prompt

GPU validation run

~1.87tok/s

CPU-only output

Low-power validation run

2systems

Independent profiles

Both validation runs passed

RLX Infra — Test Systems

Measured development validation profiles

GPU TEST SYSTEM

NVIDIA 16 GB-Class Validation Node

PASS

Measured local inference validation on a discrete NVIDIA GPU test system. Public results are shown below; the internal RLX orchestration and optimization methods used during the run are intentionally not disclosed.

Physical VRAM 15.9 GB

Local Model gpt-oss-safeguard:20b

Backend Ollama / Local

Model State Warm / keep-loaded

Measured Decode 132.7 tok/s

Measured Prompt 16,987.4 tok/s

Context Test 2,186 tokens

Measured Wall Time 16.8 s

CPU TEST SYSTEM

Low-Power CPU Validation Node

PASS

A deliberately constrained CPU-only test environment used to validate local execution while normal desktop and remote workloads remain active.

CPU Intel N5105 — 4C/4T

RAM 16 GB

Graphics Intel UHD integrated

Local Model qwen2.5-coder:1.5b

Backend Ollama

Inference CPU-only

Context Test 2,048 tokens

Measured Output ~1.87 tok/s

Cold Load ~8.7 s

Test Load Dual display + remote/screen-share active

SYSTEM A // 132.7 tok/s measured decode

SAME RLX INFRA FOUNDATION

SYSTEM B // ~1.87 tok/s CPU-only

What the Public Results Show

Measured evidence without implementation disclosure

Validation Signals

Two hardware classes Validated

Local execution Confirmed

CPU fallback path Confirmed

Accelerated inference Confirmed

Measured benchmark data Published

Adaptive behavior RLX managed

Internal methods Private

Public Diagnostics Console

Measured public output

local validation / skynet2

realystic@skynet2:~$ rlx-infra verify --public-profile
RLX Infra Measured Public Validation
   Test System A: Decode 132.7 tok/s // Prompt 16,987.4 tok/s // 2,186 tok // 16.8 s wall ... PASS
   Test System B: CPU ~1.87 tok/s // Context 2,048 // Cold load ~8.7 s .................... PASS
   Execution range: CPU-only through discrete-GPU local inference .......................... CONFIRMED
   Internal orchestration: [ PROPRIETARY / NOT DISCLOSED ]

realystic@skynet2:~$ echo "Measured results public. Internal method private."
Measured results public. Internal method private.