5G-Advanced Release 18: New Features for Mobile Network Performance

For the first three 5G releases, the story was predictable: wider channels, denser MIMO, lower latency. Release 18 breaks that pattern. Instead of chasing peak rates, 3GPP focused on making 5G intelligent, sustainable, and responsive to specialized services like XR, RedCap, and non-terrestrial networks. If you design modems, manage RAN, or evaluate enterprise connectivity, these are the features that will define the next three years.

Why Release 18 Is Called 5G-Advanced

3GPP formally rebrands Release 18 as 5G-Advanced for a reason. Previous releases optimized the data pipe; R18 optimizes the decision-making around the pipe. Two architectural shifts underpin everything:

• AI/ML as a first-class citizen: Not an overlay, but native framework for air interface, NG-RAN and core — with model lifecycle management, data collection, and fallback.
• Service-aware RAN: The base station now understands application requirements for XR, low-power IoT, and sidelink rather than treating every bearer as best-effort.

This is also a bridge. Many R18 study items — AI-based CSI compression, duplex evolution, ambient IoT — are foundation work for 6G. Understanding R18 is essential context if you follow our analysis of 6G terahertz research, where the same AI-native concepts extend to sub-THz links.

1. AI/ML for the Air Interface and NG-RAN

The most consequential change is not a higher modulation order — it is letting machine learning optimize the radio.

How it works

R18 introduces a framework for AI/ML-based CSI feedback, beam management, and positioning. Instead of a UE reporting a huge Type II codebook, an autoencoder compresses CSI at the UE into a 20-30x smaller latent vector, reconstructed by a decoder at the gNB. Both sides train collaboratively but run inference independently. 3GPP does not mandate a model — it defines data collection, model transfer, monitoring, and fallback to legacy if inference drifts.

RAN intelligence extends beyond the air interface. The NG-RAN now supports AI-enabled load balancing, mobility optimization, and energy saving with defined inputs (UE trajectory, resource status) and outputs (handover recommendation, cell activation). This standardizes what was previously vendor-proprietary SON.

For on-device context, this pairs directly with modem-side inference. Modern SoCs already run ML accelerators for CSI and beam prediction — as we detailed in our on-device AI agent analysis, the efficiency of that accelerator determines whether AI-CSI saves power or costs it.

Measured gain: Cross-vendor trials report 25-60% reduction in CSI overhead and 12-18% cell-edge throughput gain with AI-based beam prediction, especially at FR2 where beam misalignment costs are highest.

2. MIMO Evolution: Beyond 32 Ports

MIMO in R18 is about precision and uplink balance, not just larger arrays.

• Enhanced Type II codebooks for 32 ports: Support for coherent joint transmission (CJT) across multiple TRPs with tighter phase calibration for FDD massive MIMO.
• Uplink MIMO lift: Support for 8Tx UE operation (up from 4Tx), 4 layers, and SRS enhancements. Critical for uplink-heavy use cases like XR and live broadcast.
• DMRS and CSI-RS density optimization: Adaptive overhead — fewer RS resource elements when channel is static, freeing 6-8% of REs for data.
• Unified TCI framework extension: Faster beam switching (within a slot) extended to multi-TRP, reducing beam failure recovery from ~20ms to ~8ms.

Hardware implication: 8Tx uplink is less about the modem than the RF front-end. Handset OEMs must route 8 SRS ports with acceptable isolation and SAR. Expect this first in fixed-wireless routers and flagship phones that already handle complex Snapdragon 8 Gen 4-class RF architectures with discrete PAMiDs per band.

3. XR and Real-Time Media: Making the Network Application-Aware

Release 17 treated XR like faster eMBB. Release 18 fixes that. XR traffic is periodic, jitter-sensitive, and structured as PDU Sets — groups of packets where losing one PDU can make the whole set worthless (e.g., a video frame).

Key additions:

• PDU Set awareness: The core marks PDU Set importance and dependency; the RAN can discard low-importance sets intentionally under congestion instead of randomly dropping packets.
• L4S and delay status reporting: Explicit congestion notification plus UE reporting of buffer delay allows the scheduler to prioritize XR before queuing delay spikes.
• Configured Grant enhancements: Multiple CG PUSCH occasions with dynamic offset handle XR uplink jitter without grant overhead.

In practice, this cuts XR capacity requirements by 30-40% at a given satisfaction rate (95% of users with <20ms mouth-to-glass latency). If you build fixed-wireless CPE or AR glasses, PDU Set handling in the modem is now part of conformance testing.

4. Network Energy Saving and Enhanced Mobility

Energy: From symbol to system

RAN energy is ~70% of network OPEX. R18 Network Energy Saving (NES) allows cell DTX/DRX — the gNB can micro-sleep its PA and transceiver chain in 0.5-10ms gaps aligned with sparse SSB and SSB-less SCell operation. Adaptive antenna port muting dynamically reduces from 64 to 32/16 ports when load drops. Unlike previous sleep modes, NES is coordinated with UE behavior via new signaling so devices do not waste power searching for dormant signals.

Operators in early trials report 18-22% site energy savings with no coverage loss when NES is paired with intelligent carrier shutdown. For device makers, the benefit is longer DRX cycles and cleaner interference environments.

Mobility: L1/L2-Triggered Mobility (LTM)

Legacy L3 handover involves RRC reconfiguration and 40-60ms interruption. R18 LTM pre-configures target cells and triggers handover via L1 measurement + MAC CE, shrinking interruption to ~10ms and reducing failure rates in FR2. This is essential for mmWave mobility and for high-speed scenarios like UAVs and high-speed rail.

5. NTN, Sidelink, and RedCap Maturity

Three previously niche areas become commercially viable in R18:

Non-Terrestrial Networks (NTN): R17 introduced NTN, but R18 adds mobility with terrestrial networks, NTN-NTN handover, 10-30 kbps coverage enhancement, and support for VSAT/earth stations. Shared access with TN using the same NR waveform is now defined for n255/n256 bands.

Sidelink evolution: Unlicensed sidelink (SL-U at 5 GHz), sidelink carrier aggregation, and enhanced relay (UE-to-network and UE-to-UE) expand V2X and public safety use. New resource allocation mechanisms reduce collision probability by 40% in dense scenarios. Cooperative awareness and sensor sharing for autonomous systems depend on this.

RedCap and Ambient IoT: 5 Mbps RedCap gets extended DRX (up to 10,000s), wake-up receiver study, and 20 MHz operation in FR1. R18 also introduces Ambient IoT — battery-less devices powered by energy harvesting for inventory and sensor tags, with a new extremely low-complexity air interface study that will mature in Release 19.

Security, Privacy, and Implementation Realities

As a security researcher, three R18 considerations stand out that often get buried under performance graphs.

AI/ML attack surface: Model poisoning or adversarial CSI inputs could degrade beamforming for a cell. R18 mandates model monitoring, version control, and deterministic fallback to non-AI operation — but operators must validate that fallback actually triggers. Model transfer over RRC also needs integrity protection; ensure your update path is signed.

Slicing isolation: Enhanced slicing with NSACF (Network Slice Admission Control Function) improves quota enforcement, yet RAN slicing still shares physical resources. Verify isolation testing, especially when mixing enterprise and consumer slices. Review your device's handling of slice credentials — similar hygiene to what we recommend in our mobile password manager security guide applies: hardware-backed storage, not app-level secrets.

Sidelink privacy: Unlicensed sidelink and relay increase device discoverability. R18 adds privacy enhancements (temporary IDs, restricted discovery), but misconfigured relays can leak location. Audit PC5 provisioning.

Finally, power and thermal remain the quiet constraint. 8Tx SRS sounding, continuous AI inference, and SSB-less measurements are incremental — but together they move peak modem power by 300-500 mW. Thermal design for sustained XR sessions matters more than benchmark uplink numbers.

What to Build For Now

If you are planning a device or enterprise deployment for 2026–2027:

1. Require R18 modem certification for NTN or XR products. R17 NTN does not hand over to terrestrial networks reliably.
2. Test LTM mobility explicitly on FR2. Lab handovers look perfect; urban canyon beam blockage is where LTM wins.
3. Implement PDU Set marking at the application server. Without upstream marking, the RAN cannot prioritize correctly.
4. Validate NES interoperability — some early UEs keep searching during cell DTX off-durations, erasing energy gains with battery drain.

Release 18 sets the architectural pattern for the rest of 5G-Advanced and for early 6G. The winners will not be those with the widest channel, but those whose scheduler understands what the traffic actually is.