PulseAugur
EN
LIVE 12:22:58

llama.cpp MoE cache nearly doubles decode speed for large models

By PulseAugur Editorial · [1 sources] ·

The llama.cpp project has introduced a new Mixture of Experts (MoE) cache feature that significantly boosts decoding speed for large models that do not fit entirely into GPU memory. This cache stores frequently used experts in VRAM, reducing the need to transfer data from slower system RAM over PCIe. Benchmarks show a nearly twofold increase in decoding speed on a Tesla T4 GPU with a Qwen3-30B-A3B model, and even greater improvements on higher-end hardware like the RTX 5090. AI

IMPACT Improves inference efficiency for large MoE models on consumer hardware, making them more accessible.

RANK_REASON This is an optimization for an existing open-source inference engine, not a new model release or core research.

Read on dev.to — LLM tag →

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →

llama.cpp MoE cache nearly doubles decode speed for large models

How we ranked this

Signal score
28 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Tool
This is an optimization for an existing open-source inference engine, not a new model release or core research.
Source corroboration
Single-source cluster
Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics
infra, model release
Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance
High
Clearly on-topic for AI-industry coverage.
Story freshness
Breaking (< 6h)
Fresh story with cross-source coverage still developing. Ranking may shift as more sources report.

Full methodology in our editorial standards.

COVERAGE [1]

  1. dev.to — LLM tag TIER_1 English(EN) · sam ros ·

    llama.cpp's MoE cache nearly doubled 30B decode speed on a free T4

    <p>If your MoE model is too big for your GPU, llama.cpp fills VRAM with as many whole layers of experts as fit and runs the rest on the CPU, from system RAM. Every token then waits on the slowest part of your machine. Since build b11480 there's a better use for that VRAM: a cache…

RELATED ENTITIES

RELATED TOPICS