LLMs & Models Articles
Browse 131 articles about LLMs & Models.
What Is the Gemma 4 Mixture of Experts Architecture? How 26B Parameters Run Like 4B
Gemma 4's MoE model activates only 3.8B of 26B parameters at a time using 128 tiny experts. Learn how this delivers 27B-class intelligence at 4B compute cost.
What Is Gemma 4? Google's First Apache 2.0 Multimodal Reasoning Model
Gemma 4 ships under an Apache 2.0 license with native audio, vision, function calling, and reasoning. Here's what makes it a breakthrough for open-weight AI.
ARC AGI 2 vs Pencil Puzzle Bench: The Benchmarks That Expose AI Capability Gaps
These two benchmarks test reasoning you can't fake with training data. See how GPT-5.2, Claude, Gemini, and Chinese models actually compare.
What Is Benchmark Gaming in AI? Why Self-Reported Scores Are Often Inflated
Kimi K2 reported 50% on HLE but independent testing found 29.4%. Learn how benchmark gaming works and how to evaluate AI models honestly.
What Is the China AI Gap? Why Chinese Models Lag on Benchmarks That Can't Be Gamed
ARC AGI 2 and Pencil Puzzle Bench reveal Chinese frontier models score like Western models from 8 months ago. Here's what the data shows.
What Is the Frontier Math Benchmark? Why Open Research Problems Expose True AI Reasoning
Frontier Math uses unpublished problems that take researchers days to solve. Models with full Python access still score under 3%. Here's why it matters.
What Is Gemma 4's Audio Encoder? How the E2B and E4B Models Handle Speech Recognition
Gemma 4's edge models have a 50% smaller audio encoder than Gemma 3N, with 40ms frame duration for more responsive transcription. Here's how it works.
What Is Gemma 4's Mixture of Experts Architecture? How 26B Parameters Run Like a 4B Model
Gemma 4's MoE model has 128 experts with 8 active per token, giving you 27B-level intelligence at 4B compute cost. Here's the architecture explained.
Gemma 4 vs Qwen 3.6 Plus: Which Open-Weight Model Is Better for Agentic Workflows?
Gemma 4 ships with Apache 2.0 and native function calling. Qwen 3.6 Plus has a 1M token context window. Here's how they compare for agent use cases.
What Is the Humanities Last Exam Benchmark? How Independent Testing Revealed a 21-Point Score Inflation
Kimi K2 self-reported 50% on HLE. Independent testing found 29.4%. Here's how the HLE benchmark works and why third-party verification matters.
What Is Andrej Karpathy's LLM Knowledge Base Architecture? The Compiler Analogy Explained
Karpathy's LLM knowledge base treats raw articles like source code and compiles them into a queryable wiki. Here's the full architecture breakdown.
What Is the LLM Knowledge Base Index File? How Agents Navigate Without Vector Search
Karpathy's LLM wiki uses an index.md file as a navigation map so agents can find information without semantic search or vector databases.
LLM Wiki vs RAG for Internal Codebase Memory: Which Approach Should You Use?
Karpathy's wiki approach uses markdown and an index file instead of vector databases. Here's when each method works best for agent memory systems.
What Is the Pencil Puzzle Benchmark? The Test That Measures Pure Multi-Step Logical Reasoning
Pencil Puzzle Bench tests constraint satisfaction problems with no training data contamination. GPT-5.2 scores 56%. Chinese models score under 7%.
What Is the SWE-Rebench Benchmark? How Decontaminated Tests Expose Chinese Model Inflation
SWE-Rebench uses fresh GitHub tasks that models haven't seen in training. Chinese models that match Western scores on SWE-bench drop significantly here.
Gemma 4 E2B vs E4B: How to Run a Multimodal AI Model on Your Phone
Gemma 4's edge models support audio, vision, and function calling in under 4B parameters. Here's how to run them locally on Android and iOS devices.
How to Run Gemma 4 Locally on Your Phone or Laptop With the Google AI Edge Gallery
Google AI Edge Gallery lets you download and run Gemma 4 models locally on Android and iOS with no cloud connection. Here's how to set it up in minutes.
What Is Gemma 4? Google's Apache 2.0 Open-Weight Model With Native Audio and Vision
Gemma 4 ships under Apache 2.0 with native audio, vision, function calling, and thinking. Here's what makes it different from every previous Gemma release.
What Is Microsoft MAI Transcribe 1? The Speech Model That Outperforms Whisper and Gemini Flash
MAI Transcribe 1 achieves best-in-class accuracy across 25 languages and beats Whisper, Gemini Flash, and GPT Transcribe on word error rate benchmarks.
What Is Anthropic's Prompt Caching and Why Does It Affect Your Claude Subscription Limits?
Anthropic uses prompt caching to reduce compute costs. When third-party tools break caching, your session limits drain faster. Here's the technical explanation.