Miatz playground

Open-Weights Garage

Watch VRAM fill up as you pick a model size.

Coming to the Build-LabAIE-110BYOK AIAI-Coredev · data

The Demystify signature

VRAM and quantization are usually a spec-sheet abstraction; the rack visual makes 'this model doesn't fit' a literal, watchable overflow instead of an out-of-memory error the learner has to decode later.

How it works

What goes in, what comes out

What it does

Learner picks an open-weight model family and size, then a quantization level, and a literal server-rack visual fills with GB blocks in real time showing the exact VRAM math before a single token generates. The same prompt then runs against that configuration on a routed open-weight endpoint side by side with a closed-model equivalent, reporting tokens/sec, an eval-score quality delta versus fp16, and true cost-per-1M-tokens with no vendor thumb on the scale.

You bring

A prompt or small eval set, a target model size, and a quantization choice.

You get

A VRAM fill visualization with exact GB math, tokens/sec throughput, an eval-score quality delta vs. full precision, and a side-by-side cost/latency comparison against a closed model.

You control

Model family/size, quantization level (fp16/int8/int4), hosting mode (routed hosted vs. simulated self-host VRAM math), and the closed-model comparator.

Where it's used

Free top-of-funnel lead magnet (sovereignty/anti-lock-in hook) and in-module lab for AIE-110 (elective, Open-Weights badge).

open-weight model families (Llama/Mistral/Qwen/Gemma-class)VRAM mathquantization (fp16/int8/int4)tokens/sec throughputself-hosted vs. routed hosting

Optional · watch & try

Go deeper, elsewhere

Hand-picked public explainers and open tools that complement this one — always optional, never required, never graded.

What is LoRA? Low-Rank Adaptation for finetuning LLMs EXPLAINEDExplains the low-rank matrix trick that makes LoRA cheap LoRA for fine-tuning LLMs explained with exampleWorked example applying LoRA to an actual model LLM Fine Tuning Explained in 8 Minutes: LoRA, QLoRA, DPOFast comparison of fine-tuning method families · ~8 min Hugging Face PEFT — LoRA methodsOfficial reference implementation and guide for LoRA/PEFT fine-tuning LLM Model VRAM Calculator (Hugging Face Space)Estimate GPU memory needed to fine-tune or run a given model at a given quantization

Concepts click when you open the machinery.

Three labs are already live and free — the same hands-on style this playground brings to its module.

Play the free labs Take the DSAT