PulseAugur
EN
LIVE 09:52:10
Русский(RU) Qwen3.8. MTP и тензорный параллелизм в llama.cpp: 75 ток/сек на двух RTX 3090 Плотная 27-миллиардная модель на двух RTX 3090 в llama.cpp из коробки даёт 32 токе

Qwen3.8 LLM optimized for faster inference with llama.cpp

By PulseAugur Editorial · [1 sources] ·

A technical guide details how to optimize the Qwen3.8 large language model for faster inference using llama.cpp. The author explains how to leverage tensor parallelism and multi-token prediction to achieve up to 75 tokens per second on two RTX 3090 GPUs, significantly improving upon the default 32 tokens per second. The guide provides specific flags and configurations to address bottlenecks and maximize performance. AI

IMPACT Optimizing LLM inference speed with llama.cpp can accelerate local deployment and experimentation for AI developers.

RANK_REASON The article details optimization techniques for an existing LLM using a specific software framework, rather than announcing a new model or research.

Read on Mastodon — mastodon.social →

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

Qwen3.8 LLM optimized for faster inference with llama.cpp

How we ranked this

Signal score
17 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Tool
The article details optimization techniques for an existing LLM using a specific software framework, rather than announcing a new model or 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. Mastodon — mastodon.social TIER_1 Русский(RU) · [email protected] ·

    Qwen3.8. MTP and tensor parallelism in llama.cpp: 75 tokens/sec on two RTX 3090 A dense 27-billion parameter model on two RTX 3090 in llama.cpp out-of-the-box yields 32 tokens/sec

    Qwen3.8. MTP и тензорный параллелизм в llama.cpp: 75 ток/сек на двух RTX 3090 Плотная 27-миллиардная модель на двух RTX 3090 в llama.cpp из коробки даёт 32 токена в секунду. Интуитивно понятно, что это не предел: две карты по 24 ГБ и 936 ГБ/с каждая не должны выдавать столько, ск…

    LINKS habr.com/…/1076080

RELATED ENTITIES

RELATED TOPICS