We Tried to Break a 27B Q4 Model. The Benchmark Broke First.
A RAMGPT ExactBench research note on Qwen3.8-27B Q4: 94 short deterministic reasoning items, zero wrong answers, and several benchmark-design failures uncovered along the way.
Understand the systems.
Test the assumptions.
Independent experiments, source-level analysis, and practical guides to the AI running on your hardware.
Flagship independent research
Evidence & experiments
On the test bench
Adaptive Speculative Decoding on RTX 4090: Fixed n=7 Was 13.5% Faster
Independent RTX 4090 CUDA testing of adaptive DFlash2 in llama.cpp: fixed n=7 reached 80.11 tok/s on structured JSON, versus 70.61 tok/s for adaptive 3–7.
HereticBench: What Actually Changes When You Remove a Model's Refusal Behavior?
A paired benchmark of refusal behavior, calibrated reasoning, open-ended response drift, and epistemic integrity in a 27B Base vs Heretic model comparison.
DSpark PCTree on an RTX 4090: Higher Acceptance, Lower Throughput
An independent llama.cpp PCTree test on RTX 4090 with Qwen3-8B Q8: DSpark doubled decode speed, while wider trees raised acceptance but still lost throughput.
RAMGPT QuantBench
Compare the artifact. Inspect the evidence.
GGUF artifact rankings against pinned BF16 references, with versioned benchmark contracts.
A cumulative learning series
AI Foundations.
Build your understanding, one idea at a time. Beginner-friendly explanations from contributors learning AI systematically.
Explore the learning path ↗- 01↗
- 02↗
Continue learning
What Does It Mean for a Machine to Learn?
- 03↗
Continue learning
Parameters, Weights, and Biases: What Are They in a Neural Network?
- 04↗
Continue learning
What Is a Tensor? The Shape of Data Inside AI
- 05↗
Continue learning
What Is a Forward Pass? How Data Moves Through a Neural Network
- 06↗
Continue learning
How Does a Neural Network Start Learning? Training Begins With Loss
- 07↗
- 08↗
- 09↗
Continue learning
Embeddings: How AI Turns Meaning Into Coordinates
- 10↗
Continue learning
Tokenization: How Text Becomes the IDs an AI Model Can Read
- 11↗
Continue learning
Attention: How Each Token Decides Which Other Tokens Matter
- 12↗
Continue learning
Transformer: How Attention Becomes a Complete Neural Network Block
- 13↗
- 14↗
From understanding to practice
Run it. Build it. Work it out.
How Much VRAM Do You Need for Local LLMs?
Learn how to estimate LLM VRAM requirements from model size, quantization, KV cache, context length, runtime overhead, and GPU offloading.
llama.cpp for Beginners: Build, Run and Tune Your First Local LLM
A simple step-by-step llama.cpp guide for beginners: compile from source, run a GGUF model, use GPU acceleration, and tune the main inference settings.
Why Is My Local LLM Slow? Common Causes and Fixes
Learn why a local LLM runs slowly and how model size, GPU offloading, context length, memory bandwidth, CPU threads, and runtime settings affect speed.
Recent writing
The complete archive ↗Fine-Tuning: How a Pretrained Model Learns a New Job Without Starting Over
AI Foundations #14 explains how fine-tuning continues gradient-based training from pretrained weights, using narrower data to change model behavior without relearning language from scratch.
vLLM Mistral-Large-3 YaRN Regression: Why Accuracy Fell 4.58 GSM8K Points
A Sep 18 vLLM fix traces a silent Mistral-Large-3 accuracy regression to config remapping: YaRN frequency interpolation was preserved, but its no-mscale flag was not reaching DeepSeek-style attention.
llama.cpp Vulkan Wrong Results: The Sliced KV Cache Regression
A Sep 15 llama.cpp Vulkan fix targets silent wrong results when mul_mat reads a slice of a larger cache, a failure more dangerous than a clean crash.
Mining Enterprise CVE Remediation Paths from Operational Tickets
A practical research playbook for reconstructing CVE-centric remediation event logs from change, incident, request, and vulnerability records before forming hypotheses.
Pretraining: How a Random Transformer Learns From Next-Token Prediction
AI Foundations #13 connects the Transformer architecture to the training objective that turns random parameters into a useful language model.
llama.cpp Errors and Fixes: A Practical Troubleshooting Guide
A source-backed llama.cpp troubleshooting map for unknown architectures, wrong tensor shapes, corrupted GGUFs, CUDA OOM, multimodal mistakes, and runtime regressions.
The people behind the work
Independent voices. Shared curiosity.
Technical Research & Editorial Team
RAMGPT Editorial Team
Local AI Systems & Inference Contributor
bettercallcaleb
AI Fundamentals Contributor
William
Open-Source AI Commentary Contributor
MapleKernel
Student Research Contributor — Mathematics & AI
machaMochaLatte
Platform Engineering & MLOps Contributor
lanceryan111
Enterprise Security Research Contributor
Lisa S
Infrastructure support
Compute behind the research
Independent AI testing can require substantial compute. We acknowledge infrastructure providers that help RAMGPT run larger experiments at practical cost.
Editorial independence
Disclosure details →Infrastructure relationships do not determine RAMGPT's benchmark methodology, measurements, rankings, or editorial conclusions.