Field guide
Education
Everything on this site — the leaderboard, the axes, the round writeups — assumes some vocabulary: quants, MoE, stalls, judges. These guides build that vocabulary from zero, in plain language, using this site’s own data as the examples. No hand-waving, no invented numbers.
How Local AI Works
The concepts cluster: what a local model actually is, why one model ships as a dozen different files, and why some models think out loud — or think forever.
What Is a Local LLM, and Why Would You Run One?
A local model is a file on your disk and math on your own silicon. What that buys you — privacy, zero marginal cost, and total version control — and what it actually takes to run one.
Part 2Model Size and Quantization: Why Smaller Isn't Always Worse
What parameters and bit-depths actually mean, how to read Q4_K_M vs Q6_K vs 8bit, and the round where a mid-size quant beat the biggest one outright.
Part 3GGUF, MLX, Safetensors: Why One Model Ships in a Dozen Packages
The same model gets converted, quantized, and repacked by different hands for different engines — and the packaging can change behavior, not just speed. What each format is and why provenance matters.
Part 4Dense vs. Mixture-of-Experts: How a 26B Model Runs Like a 4B
MoE models activate only a slice of their parameters per token — which is why a 26B can generate five times faster than a 27B on the same machine. What the A-number means, and the memory catch.
Part 5Speculative Decoding and MTP: Faster Without Getting Dumber
Draft-and-verify generation can nearly double a model's speed with identical output quality — but only when the model and engine both support it. Why 'which runtime is faster' has become a per-model question.
Part 6Reasoning Models: Chain-of-Thought, Effort Knobs, and the Stall
Thinking models write a private scratchpad before answering — usually helpful, sometimes catastrophic. What reasoning_effort actually does, and the failure mode where a model burns 16,000 tokens and says nothing.
How GAUNTLET Scores Models
The methodology cluster: what the eight axes measure, what a test actually looks like, how to read a score honestly — including the ways scores can mislead — and why the same model can score differently on different days.
The Eight Axes, Explained Like You're New Here
G-A-U-N-T-L-E-T: eight letters, eight different questions about a model. A beginner's tour of what each axis measures, why one is counted rather than judged, and how to read a radar chart honestly.
Part 8Inside a Test Suite: What We Ask, How We Judge, What Stays Private
A walk through an actual GAUNTLET test using the two public community prompts we run verbatim — plus how 0–20 judging works, why temperature 0, and the contamination policy in plain words.
Part 9Reading a Score: What a Zero Means, and When It Doesn't
A 0/20 can be a failed model, a stalled model, or a broken judge — and we've published examples of all three. How to read GAUNTLET numbers with the right amount of trust.
Part 10Why Scores Differ Day to Day: A Field Guide to Confounds
Same model, same suite, different number — how quants, engine versions, effort settings, prompt design, and run order each move benchmark results, and what we do about each one.
Where to go next
For the formal scoring rules, see the methodology page. For the evidence behind every number cited in these guides, see the round writeups. And the leaderboard is where it all lands.