885 related articles

Why memory bandwidth (GB/s), not VRAM size, determines local LLM inference speed. Includes tokens/sec formula, GPU bandwidth comparison, and a practical card selection framework.

Learn why memory bandwidth (GB/s)—not VRAM size—determines local LLM inference speed. Get the tokens/sec formula, GPU bandwidth comparisons, and a practical card selection hierarchy.

In-depth analysis of Apple Silicon local LLM inference speed benchmarks covering M-series memory bandwidth, model quantization, MLX framework optimization, and Mac configuration guidance.

D-Flash uses fast diffusion parallel drafting and target feature KV injection to solve the autoregressive Drafter latency bottleneck in speculative decoding. 16 tokens in just 6ms, up to 3.5x speedup on HumanEval, surpassing EAGLE3 and MTP.

DeepSeek and Peking University open-source DSpark, an inference acceleration technology using semi-autoregressive architecture and dynamic scheduling to boost LLM speed by 50%+ and double GPU concurrency without quality loss.

DeepSeek open-sources DSpark, delivering 50%–400% LLM inference speedup with no retraining or quantization, via semi-autoregressive drafting and confidence-scheduled verification.

DeepSeek V4 Flash launches with benchmark scores approaching Claude Opus 4.8 at just $0.18 per million output tokens. Deep analysis of performance, pricing, and industry impact.

A complete guide to building a local private AI assistant with Ollama and Qwen-Agent. Covers RAG knowledge integration, voice interaction, and permission isolation for a secure local AI Agent architecture.

A developer proposes a Flex API-based slow mode for Codex, trading speed for nearly double the usage quota. We analyze the product logic, technical feasibility, and business challenges.

Deep analysis of the real cost of serving a 2.8 trillion parameter model. From MoE sparse activation to batching scale effects and inference optimization, revealing why model size and serving cost are less correlated than assumed.

As AI LLM capabilities converge, cost-effectiveness becomes the key selection factor. This article explores how to rationally compare AI models through value assessment, task matching, and cost-benefit analysis.

Deep dive into Heretic uncensoring technology applied to Jamba2-Mini, Qwen3.5-9B, and 27B open-source models, exploring how refusal rates dropped from 97% to 4% and the safety debates involved.

An in-depth analysis of why teams are abandoning LLM routers, exploring hidden complexity costs, outdated cost assumptions, and how to avoid over-engineering in AI systems.

Orca-Bench is a benchmark for evaluating AI agents' operational capabilities, testing LLMs on fault diagnosis, multi-tool orchestration, and risk decisions in simulated Oncall scenarios.

Complete guide to setting up a local AI coding environment on MacBook Pro M4, covering Ollama, MLX, Continue, Qwen3-Coder 30B configuration, and performance optimization strategies for 32GB RAM.

A systematic learning path for understanding the Kimi K3 technical report, covering MoE, MLA, distributed training, and modern post-training techniques.

GPT-5.6 Luna tops Google's flagship on the Artificial Analysis Intelligence Index while priced below Google's entry-level model. A deep dive into what this performance-cost breakthrough means.

DeepSeek V4 Flash model weights reportedly open-sourced. This article analyzes its lightweight positioning, open-weight value, comparisons with closed-source models, and deployment guidance.

Running Kimi K3 with 29GB RAM at just 0.5 tok/s. An in-depth analysis of extreme quantization techniques, performance trade-offs, and the impossible triangle of local LLM deployment.

Running Kimi K3 with 29GB RAM at just 0.5 tok/s. A deep analysis of extreme quantization techniques, performance trade-offs, and the impossible triangle of local LLM deployment.