Research

October 2026Retrieval & Quantization Research

One Fact per Vector, and What Retrieval Does to a Quantized Reader

July 2026Implementation

How to Screen a Compressed Model for Agent Safety Before You Ship It

July 2026Production AI

Your Compressed Models Pass Every Benchmark. That Is Not the Same as Agent-Safe

July 2026AI Safety

Fidelity Is Not Safety

June 2026Evaluation Methodology

The General-Knowledge Floor: Why Knowledge-Injection Evals Cluster at 70%

May 2026MoE Research

MoE Routing Layers Converge Across Subjects: No Free Lunch for Domain-Specific Targeting

May 2026MoE Research

Metal Buffer Limits Block LoRA Scaling on MoE Models

May 2026MoE Research

The Stacking Confound: Why LoRA Recovery Numbers Lie

May 2026MoE Research

Identity Survives LoRA Stacking: Persona Preservation in Multi-Stage Fine-Tuning

May 2026Reproducibility

Decision Gates as Forcing Functions: Green / Yellow / Red Pre-Registered Verdicts

May 2026Reproducibility

Pre-Registering Predictions in ML Research: A Worked Example

May 2026LoRA Engineering

Multimodal LoRA on Text-Only Data: The Visual-Block Gradient Sink

May 2026LoRA Engineering

PEFT `target_modules` Has Three Modes, Most Teams Use the Wrong One

May 2026Datasets

OpenThoughts-114k Was Removed From the Hub. Here's the Substitute.

May 2026Eval Methodology

KLD Tail Effects: Why Quality Comparisons Need Median, Not Mean

May 2026Eval Methodology

MATH-500 Boxed-Answer Extraction: Five Edge Cases That Tank Your Score

May 2026Eval Methodology

Why Your Single-Seed GPQA-Diamond Comparison Is Probably Noise

May 2026MoE Research

MoE Routing Layers Converge Across Subjects: No Free Lunch for Domain-Specific Targeting

May 2026Quantization

Quantizing a Trillion-Parameter MoE in 15 GB of RAM

May 2026Quantization

When Quantizing FP8 Models Produces Garbage: The Block-Scale Trap

May 2026MoE Research

Metal Buffer Limits Block LoRA Scaling on MoE Models

May 2026Quantization

Hybrid SSM-Attention Models Lose 7 pp When Converted to GGUF

May 2026MoE Research

The Stacking Confound: Why LoRA Recovery Numbers Lie

May 2026MoE Research

Identity Survives LoRA Stacking: Persona Preservation in Multi-Stage Fine-Tuning

May 2026Apple Silicon Engineering

A 16 GB Mac Mini Can Quantize a 250 GB Model

April 2026Research Report

Gemma 4 at Half the Size, But Full Performance

March 2026MoE Research

What 100 Prompts Reveal About Expert Routing in 256-Expert MoE Models

March 2026MoE Research

Why You Can’t Prune MoE Experts, Even the Ones Nobody Uses

March 2026RAM Research

SmoothQuant Breaks on Signed RMSNorm: A Negative Result Worth Knowing

March 2026RAM Research

Why a Mostly-6-Bit Model Runs Faster Than a Mixed 4/8-Bit Model

March 2026RAM Research

When Better Is Worse: What We Learned by Trying to Improve Our Quantizer

March 2026RAM Research

Five Things Our Benchmarks Reveal That Nobody Expected

March 2026RAM Research

RAM Benchmark Results: 7 Models, 40,000+ Questions, One Winner

March 2026RAM Research

Beyond Perplexity: Downstream Benchmarks Confirm RAM Beats All Quantization Strategies

March 2026RAM Research

Does Quantization Actually Regularize? We Tested It.

March 2026RAM Research

Mean Perplexity Is Lying to You

March 2026RAM Research

MLX Quantization on Apple Silicon: How RAM Turns a Mac into a Model Compression Lab

March 2026RAM Research

When Data-Free Beats the Gold Standard

March 2026RAM Research

The GPU Hours Nobody Needed to Spend

March 2026RAM Research

The Quantization Bottleneck Is About to Break

March 2026Research

What RAM Actually Delivers: Evidence from Four Models and 20,000 Tensors

March 2026Evaluation Report

RAM Evaluation Results: Four Models, Three Architectures, One Framework

March 2026Research Insight

Why RAM Matters: Proprietary Compression and the Future of Model Deployment

March 2026Research Insight

When Quantization Beats Full Precision: Anatomy of a Perplexity Anomaly

February 2026Privacy & AI Sovereignty

AI Without Permission: Privacy, Sovereignty, and the Case for Local Inference

February 2026MLOps & Infrastructure

The End of Calibration Data: Why RAM Makes an Entire ML Pipeline Step Obsolete

February 2026AI to Production

AI Sovereignty on Commodity Hardware: How RAM Breaks the GPU Cartel

February 2026Embodied AI

RAM and the Humanoid Intelligence Problem: Why Quantized Models Are the Brain Robotics Needs

February 2026Enterprise AI

RAM for Enterprise: Deploying Frontier AI Without the GPU Bill

February 2026Apple Silicon

RAM on Apple Silicon: Running 400B Parameter Models on a Single Mac

February 2026Quality Assurance

Why Collapse Tests Are Insufficient for Quantization Quality Assessment

February 2026Apple Silicon

MLX Quantization on Apple Silicon, Engineering Pitfalls and Workarounds

February 2026Model Optimization

Expert Pruning in MoE Models, When Dead Experts Aren't Dead

February 2026Quantization

Per-Expert Mixed-Bit Quantization via Mask-and-Combine Dispatch

February 2026MoE Research

Gemma 4 at Half the Size, But Full Performance

One Fact per Vector, and What Retrieval Does to a Quantized Reader

How to Screen a Compressed Model for Agent Safety Before You Ship It

Your Compressed Models Pass Every Benchmark. That Is Not the Same as Agent-Safe

Fidelity Is Not Safety

One Fact per Vector, and What Retrieval Does to a Quantized Reader

How to Screen a Compressed Model for Agent Safety Before You Ship It

Your Compressed Models Pass Every Benchmark. That Is Not the Same as Agent-Safe

Fidelity Is Not Safety

The General-Knowledge Floor: Why Knowledge-Injection Evals Cluster at 70%

MoE Routing Layers Converge Across Subjects: No Free Lunch for Domain-Specific Targeting

Metal Buffer Limits Block LoRA Scaling on MoE Models

The Stacking Confound: Why LoRA Recovery Numbers Lie

Identity Survives LoRA Stacking: Persona Preservation in Multi-Stage Fine-Tuning

Decision Gates as Forcing Functions: Green / Yellow / Red Pre-Registered Verdicts

Pre-Registering Predictions in ML Research: A Worked Example

Multimodal LoRA on Text-Only Data: The Visual-Block Gradient Sink

PEFT `target_modules` Has Three Modes, Most Teams Use the Wrong One

OpenThoughts-114k Was Removed From the Hub. Here's the Substitute.

KLD Tail Effects: Why Quality Comparisons Need Median, Not Mean

MATH-500 Boxed-Answer Extraction: Five Edge Cases That Tank Your Score

Why Your Single-Seed GPQA-Diamond Comparison Is Probably Noise

MoE Routing Layers Converge Across Subjects: No Free Lunch for Domain-Specific Targeting

Quantizing a Trillion-Parameter MoE in 15 GB of RAM

When Quantizing FP8 Models Produces Garbage: The Block-Scale Trap

Metal Buffer Limits Block LoRA Scaling on MoE Models

Hybrid SSM-Attention Models Lose 7 pp When Converted to GGUF

The Stacking Confound: Why LoRA Recovery Numbers Lie

Identity Survives LoRA Stacking: Persona Preservation in Multi-Stage Fine-Tuning

A 16 GB Mac Mini Can Quantize a 250 GB Model

Gemma 4 at Half the Size, But Full Performance

What 100 Prompts Reveal About Expert Routing in 256-Expert MoE Models

Why You Can’t Prune MoE Experts, Even the Ones Nobody Uses

SmoothQuant Breaks on Signed RMSNorm: A Negative Result Worth Knowing

Why a Mostly-6-Bit Model Runs Faster Than a Mixed 4/8-Bit Model

When Better Is Worse: What We Learned by Trying to Improve Our Quantizer

Five Things Our Benchmarks Reveal That Nobody Expected

RAM Benchmark Results: 7 Models, 40,000+ Questions, One Winner

Beyond Perplexity: Downstream Benchmarks Confirm RAM Beats All Quantization Strategies

Does Quantization Actually Regularize? We Tested It.

Mean Perplexity Is Lying to You

MLX Quantization on Apple Silicon: How RAM Turns a Mac into a Model Compression Lab

When Data-Free Beats the Gold Standard

The GPU Hours Nobody Needed to Spend

The Quantization Bottleneck Is About to Break

What RAM Actually Delivers: Evidence from Four Models and 20,000 Tensors

RAM Evaluation Results: Four Models, Three Architectures, One Framework

Why RAM Matters: Proprietary Compression and the Future of Model Deployment

When Quantization Beats Full Precision: Anatomy of a Perplexity Anomaly

AI Without Permission: Privacy, Sovereignty, and the Case for Local Inference

The End of Calibration Data: Why RAM Makes an Entire ML Pipeline Step Obsolete

AI Sovereignty on Commodity Hardware: How RAM Breaks the GPU Cartel

RAM and the Humanoid Intelligence Problem: Why Quantized Models Are the Brain Robotics Needs

RAM for Enterprise: Deploying Frontier AI Without the GPU Bill

RAM on Apple Silicon: Running 400B Parameter Models on a Single Mac

Why Collapse Tests Are Insufficient for Quantization Quality Assessment

MLX Quantization on Apple Silicon, Engineering Pitfalls and Workarounds

Expert Pruning in MoE Models, When Dead Experts Aren't Dead

Per-Expert Mixed-Bit Quantization via Mask-and-Combine Dispatch

Profiling Expert Activation Patterns in 512-Expert MoE Models