Zimmer picks the right model for the machine you have, keeps it current as better ones ship, and handles everything underneath. You never touch a config file.
Pick your memory and see what fits. No sign-up, no download required to find out.
Your machine's memory
3B class
Quick questions, drafting, summarizing
7–8B class
Everyday chat and writing
14B class
Stronger reasoning and coding
32B class
Serious coding and long documents
70B class
Frontier-grade reasoning
120B+ class
The largest open models available
Estimates use the same memory math the app runs on your actual machine, so what you see here is what Zimmer will tell you after you install it. Real numbers depend on the specific model file and how much context you ask for.
Auto compression sizing: Zimmer selects the optimal quantization (Q4/Q8/FP16) so models run at max speed within your exact memory budget.
Metal MLX & GGUF runtime: Native Apple Silicon MLX acceleration for Mac + cross-platform GGUF quantization for Windows.
VRAM & KV-slot budgeting: Automatically offloads GPU layers and scales KV-cache slots so models load reliably without crashing.
Multi-node mesh compute: Add Apple Silicon Macs to a shared pool with a one-time code, then route every request to whichever node is least busy and has the model loaded.
Zero configuration: No Python setups or config files. Pick a model and it installs, tunes, and runs in one click.
Upgrades find you: Zimmer checks daily whether a model you've installed has shipped a newer build, and tells you when one has.
Swapped safely: An update downloads and verifies completely before anything is replaced. A failed update leaves your working model untouched.
Nothing disappears: No provider can deprecate, restrict, or reprice the model your workflow depends on. It's on your disk.
Fast by default: A small model answers everyday questions instantly, so nobody waits on a heavyweight for a one-line reply.
Big when it matters: Long documents, research questions, and agent work engage a larger model automatically — or whenever someone asks for it.
No black box: Every routing decision is visible, with the reason it was made. You can always see which model answered.
Add another machine to the pool and requests spread across both automatically. See how teams deploy it
The whole open ecosystem: Thousands of open models, not a shortlist someone negotiated on your behalf.
The files are yours: Downloaded models sit on your disk. They keep working with or without Zimmer, online or offline.
Different models for different jobs: Run a coding model, a writing model, and a reasoning model side by side. Switch whenever you like.
Formats and compatibility
Any GGUF model from Hugging Face, including split multi-part repositories, running on the llama.cpp runtime.
MLX models run natively on Apple silicon, using the unified memory architecture directly.
Zimmer reads each file's quantization and parameter count, then picks the variant that fits your available memory.
Zimmer detects your hardware on first launch and recommends a model that will actually run well on it.