Dyno v0.2.0

Dyno 0.2.0 turns the Mac app into an AI safety and alignment research workbench, with local MLX inference as its foundation.

Research workflows

• Inspect activations in an already-serving model without loading another weight copy. See next-token probabilities and numeric activation norms on a shared color scale.
• Run isolated interventions, labeled linear probes and small SAE experiments, with held-out metrics and control baselines.
• Save activation captures and token analyses automatically. Reopen results and settings; rerunning creates a new record. Isolated job history also restores configuration.
• Choose Thinking: Default, On or Off beside Start, with per-request controls in Chat and Token analysis. Requires a compatible model template; raw-text research does not require thinking.
• Use the Python SDK, HTTP research APIs and eight local MCP tools. Documentation includes resident capture and resource/lifecycle constraints.

Reliability and model management

• Preserve hybrid decoder metadata when tapping activations (fixes Qwen is_linear failures).
• Detect already-running Dyno endpoints and expose Stop controls, port selection and clearer resource checks.
• Resolve actual model paths when an endpoint advertises default_model.
• Keep native execution history bounded; explain skipped traces separately from inference failures and release history slots even when final response parsing fails.

Install and upgrade

Download Dyno-0.2.0-arm64.dmg, open it, and drag Dyno to Applications. Requires Apple Silicon and macOS 14+. Python, MLX and the MCP launcher are bundled; model weights are separate. A SHA-256 checksum accompanies the DMG.

Existing inference processes need an explicit stop/start with this version to gain resident capture. Dyno does not automatically restart external servers. Check active requests before stopping. Previous unsaved session-only results cannot be recovered after closing the old app.

The app is ad-hoc signed, not Apple-notarized. See Apple's first-launch guidance.

These are experimental research tools, not safety certification. Resident capture returns activation norms, not individual-neuron interpretations or causal explanations. Probes, interventions and SAE training use separate workers and require adequate memory and quiet GPU capacity. SDK/MCP jobs and native app histories use distinct stores documented in the guide. LAN inference has no authentication; enable sharing only on trusted networks. Research and execution endpoints remain local-only.

Dyno v0.1.0

Run local MLX models on Apple Silicon, measure their inference speed, and share them with other computers on your network.

• Native Models, Router, Inspect, Performance, and Discover views, plus Chat and menu bar telemetry.
• LAN sharing with automatically detected addresses, a copyable URL, and a test request.
• OpenAI-compatible inference, streaming, and automatic or explicit model selection.
• Router settings and request history remain local-only.

Install

Download Dyno-0.1.0-arm64.dmg, open it, and drag Dyno to Applications. Requires macOS 14 or later on Apple Silicon. Python and MLX are bundled; model weights are downloaded separately.

The app is ad-hoc signed and not Apple-notarized. macOS may block the first launch; see Apple's guidance before choosing whether to open it.

Click the menu bar icon to open Dyno; right-click it for a quick summary. To share a running model, enable Router → Share on local network, then click Start the router. Use only trusted networks: inference is available without authentication while sharing is enabled.

A SHA-256 checksum is included alongside the DMG.