Train 600M MoE, Dense, Hybrid LLM Architectures.
The Big LLM Architecture Comparison
From DeepSeek-V3 to Kimi K2: A Look At Modern LLM Architecture Design
LLM Architecture Gallery
A gallery that collects architecture figures from The Big LLM Architecture Comparison and related articles, with fact sheets and links back to the original s...
Benchmarks | EXO
Transparent benchmarks for LLMs tested on real hardware. Coming soon.
Olmo Hybrid and future LLM architectures
The latest Olmo model and discussions at the frontier of open-source post training tools.
Kimi K2: 1 T‑Param DeepSeek‑Inspired MoE, Re‑engineered & Trained from Scratch
Still expecting Llama 4 Behemoth? Check Kimi K2.
One year as an AI Engineer: The 5 biggest misconceptions about LLM reliability I've encountered
535 votes, 59 comments. After spending a year building evaluation frameworks and debugging production LLM systems, I've noticed the same…
apple-silicon-llm-bench/results/complete_results.html at main · AlexHiesch/apple-silicon-llm-bench
Systematic LLM inference benchmark for Apple Silicon: 8 backends, 7 models, 791 measurements - AlexHiesch/apple-silicon-llm-bench
Exploring LLMs with MLX and the Neural Accelerators in the M5 GPU
Mac with Apple silicon is increasingly popular among AI developers and researchers interested in using their Mac to experiment with the…
Alex Cheema on Twitter / X
.@karpathy shouted out my work on @exolabs at @ycombinator AI SUS!“we use LLMs similarly to mainframes in the ‘70s - compute is timeshared by having a slice in the batch dimension. models will compress over time, and with this we’ll be able to run more on-device” pic.twitter.com/UKSrquQGAL— Alex Cheema (@alexocheema) June 18, 2025
Doriandarko/MLX-GRPO
A pure MLX-based training pipeline for fine-tuning LLMs using GRPO on Apple Silicon.
Megakernel: Matching Apple Silicon Efficiency at 2x the Throughput on a RTX 3090
The first megakernel for hybrid DeltaNet/Attention LLMs. 413 tok/s at 1.87 tok/J on a 2020 RTX 3090, matching M5 Max efficiency at 1.8x throughput.
Daniel Han on Twitter / X
OpenAI's OSS model possible breakdown:1. 120B MoE 5B active + 20B text only2. Trained with Float4 maybe Blackwell chips3. SwiGLU clip (-7,7) like ReLU64. 128K context via YaRN from 4K5. Sliding window 128 + attention sinks6. Llama/Mixtral arch + biasesDetails:1. 120B MoE… https://t.co/bMFp3Z6Gs5 pic.twitter.com/1NFO4utPqr— Daniel Han (@danielhanchen) August 1, 2025
Introducing LFM2: The Fastest On-Device Foundation Models on the Market | Liquid AI
Today, we release LFM2, a new class of Liquid Foundation Models (LFMs) that sets a new standard in quality, speed, and memory efficiency for on-device deployment. Built on a hybrid architecture, LFM2 delivers 200% faster decode and prefill performance than Qwen3 and Gemma 3 on CPU. It also significantly outperforms models in each size class on instruction-following and function calling—the core capabilities that make LLMs reliable for building AI agents.
Feature-Driven Architecture: Designing Scalable Applications
Introduction In the first article of this series, we discussed Atomic Design as a method...
XiongjieDai/GPU-Benchmarks-on-LLM-Inference
Multiple NVIDIA GPUs or Apple Silicon for Large Language Model Inference?