Llama 4 Scout just dropped. Run it locally today.
Run 109B Models on a 16GB Laptop
The first MoE expert surgery tool. Compress any LLM to fit your hardware.
Profiling model... MoE detected (16 experts, top-2)AI Route: expert_prune(8/16) + Q4_K_MPruning experts: 16 → 8 (removing lowest-impact)Quantizing: FP16 → Q4_K_MConverting to GGUF...Done! 218GB → 15.8GB (13.8x compression)Quality retained: 82.1% (MMLU) | 28.3 tok/s on RTX 4090→ ollama run smelt/llama4-scout-109b-q4kmParameter frontier
Consumer GPU VRAM
Fragmented ecosystems
Running uncompressed
The Problem
The Gap Is Growing
Frontier models scale from 70B to 671B parameters. Consumer GPUs are stuck at 24GB VRAM. The tools between them are fragmented across 6 incompatible ecosystems.
Model Size
70B → 405B → 671B. Each generation demands more RAM than any consumer card provides.
GPU VRAM
RTX 4090 shipped at 24GB in 2022. RTX 5090 ships at 32GB. Models grew 9x in the same window.
Tooling
GGUF, GPTQ, AWQ, EXL2, ONNX, TensorRT. Each has its own tools, formats, and quality tradeoffs.
How It Works
Three Steps to Compression
Profile
Tell Smelt your model and target hardware. It analyzes architecture, parameter count, and MoE topology.
Route
AI selects the optimal compression pipeline — quantization level, pruning targets, expert cuts — calibrated to your VRAM.
Compress
One command. Output: a GGUF file that runs on Ollama, llama.cpp, or LM Studio. Quality report included.
Try it now
What can you run on your hardware?
Select your hardware and use case. We'll show you the best open-source models that fit — ranked by quality.
No models fit 16 GB for this task. Try more RAM.
Features
Everything You Need to Compress
Six compression techniques in one pipeline. From basic quantization to advanced MoE surgery.
Proof
Compression Quality Report
Every compression job produces a quality report with before/after benchmarks. No guessing — you see exactly what you get.
Llama 4 Scout 109B compressed to 16GB GGUF with 82% quality retained
| Input | Llama 4 Scout 109B |
| Architecture | MoE (16 experts, top-2) |
| Original Size | 218 GB |
| Pipeline | expert_prune(8/16) + Q4_K_M |
| Output Size | 15.8 GB |
| Format | GGUF (Q4_K_M) |
| Compression | 13.8x |
| Quality (MMLU) | 82.1% retained |
| Perplexity | 5.42 → 6.18 (+14%) |
| Throughput | 28.3 tok/s on RTX 4090 |
Pricing
Start Free, Scale When Ready
Free
- Quantization (all GGUF formats)
- AI Route Selector
- 1 cloud job / month
- Community support
Pro
- Everything in Free
- MoE Expert Surgery
- Depth + Width Pruning
- 5 cloud jobs / month
- Priority queue
- Quality reports with benchmarks
Team
- Everything in Pro
- 20 cloud jobs / month
- Team workspace
- Custom compression recipes
- Dedicated support
- API access
Why Smelt
The Pain Is Real
109B models need 218GB RAM — you have 16GB
Zero production tools exist for MoE expert pruning
Trial-and-error quantization wastes hours for every new model
Run models nobody else can fit on your GPU
The first MoE expert surgery tool. Compress any LLM to fit your hardware.