Inspect & probe
Token-by-layer activation maps, intermediate predictions, and linear probes with held-out metrics and control baselines.
ACTIVATIONS / PROBES
Dyno Lab
Preparing the model universe…
Explore the lab now ↓
AI SAFETY & ALIGNMENT
Your AI
research lab.
Understand model behavior.
Test what changes it.
01 — THE RESEARCH WORKBENCH
Ask a question.
Run an experiment.
Investigate how models behave. Explore activations, test hypotheses, and compare interventions in one AI safety and alignment research workbench. Local MLX inference powers the experiments; the native app, Python SDK and APIs put you in control.
InspectActivations & execution tracesExperimentProbes, interventions & SAEsExtendPython SDK & research APIsRun locallyPowered by MLX on your Mac
RESEARCH LAB / FROM OBSERVATION TO EXPERIMENT
Look inside.
Test the difference.
Capture layer activations. Train a labeled probe. Patch or steer a representation. Compare the result with an unchanged baseline. Resident activation capture reuses loaded weights. Probes, interventions and SAE training run in isolated workers. Results stay on your Mac.
Change & compare
Scale, ablate, patch or steer. Compare target-token probabilities and generated answers across a strength sweep.
INTERVENTIONS / EVIDENCE
Explore sparse features
A small SAE sandbox with training curves, reconstruction metrics and activating examples. Feature meanings remain hypotheses to test.
SPARSE AUTOENCODERS
Observe a prediction. Change a representation. Test a probe. Explore sparse features. These are real saved experiments rendered in Dyno, not a live generation recording. The small sentiment dataset demonstrates the workflow; it is not a safety benchmark.
Save the evidence. Activation captures and token analyses now save automatically, including their settings and results. Reopen a run, adjust a hypothesis, and create a new experiment. Extend the workflow through Python, HTTP, or eight local MCP tools.
Research tools are experimental. Interpretability readouts do not certify safety or alignment. Start with a small supported model.
THE FOUNDATION / DISCOVER
Choose your
research model.
Explore MLX models on Hugging Face without leaving Dyno. Compare size and precision, see what fits your Mac, and download in one click.
THE OPEN MODEL ECOSYSTEM, ON YOUR DESKTOPTHE FOUNDATION / RUN
Local inference.
Your lab’s engine.
Start a model and open Chat. Tune generation, prompt caching, and concurrency. Your OpenAI-compatible endpoint is ready for the tools you already use.
A NATIVE APP. NO TERMINAL REQUIRED.THE FOUNDATION / OBSERVE
See the work.
Not just the words.
Tokens per second. Time to first token. Prompt-cache hits. Read measurements from inside the server, alongside GPU, memory, bandwidth, and power.
MEASURED THROUGHPUT. VISIBLE HARDWARE.
DynoNATIVE / macOS
Actual app captures. Measurements vary with model and workload.
02 — MAKE THE INVISIBLE VISIBLE
Every token has a story.
Read the whole thing.
Throughput
Decode speed measured inside the token stream, for streaming and non-streaming requests alike.
tokens / second
Responsiveness
Time to first token and prefill throughput reveal how long the model takes to start answering.
time to first token
Headroom
Watch unified memory, power, and GPU load together. Understand the machine behind the model.
GPU · memory · watts
03 — CONNECT YOUR WORLD
One Mac.
More possibilities.
Your models can serve more than one screen. Turn on local network sharing and connect Windows, Linux, or another Mac to the same inference endpoint.
Your Mac LOCAL INFERENCE
WindowsYOUR CLIENT
LinuxYOUR TOOLS
Another MacYOUR NETWORK
01
Start a model
Models → Start
02
Enable sharing
Router → Share → Start
03
Connect a client
Copy URL · model: auto
Sharing is off by default. Use a trusted network: devices that can reach the endpoint can submit inference without an API key. Router settings and request history remain local-only.
Connection guide & Windows / WSL examples ↗
FinderFileEdit
Dyno’s menu bar controls, captured from the app.
READY WHEN YOU ARE
Your next question.
Starts here.
Open your AI research lab. Inspect, probe and experiment.
Python and MLX are bundled. Add a model to get started.
Installing Dyno +
Download the Apple Silicon DMG from GitHub Releases, open it, and drag Dyno to Applications. Models are downloaded separately from Discover.
The app is ad-hoc signed and not Apple-notarized. macOS may block its first launch. Read Apple’s guidance for opening an unidentified developer’s app before deciding whether to open it. Each release includes a SHA-256 checksum.
Build it yourself +
Install Xcode’s command line tools and uv, then:
git clone https://github.com/canivel/dynolab
cd dynolab/app
./build.sh
For a distributable disk image, run
./package-dmg.sh on Apple Silicon.
Prefer the command line? +
uv tool install 'mlx-dyno[serve]'
dyno serve --model mlx-community/Qwen3-8B-4bit
dyno topServe models, inspect token probabilities, benchmark, and watch hardware telemetry from your terminal.