Antigravity 2.0 desktop + Gemma 4 speedups on Apple Silicon — Dev Signal
Antigravity 2.0 desktop + Gemma 4 speedups on Apple Silicon
Share:
Tool of the Week
Google ships Antigravity 2.0 agent desktop app
Standalone agent-first desktop app decouples from IDE; adds dynamic subagents, async task management, JSON hooks, and scheduled execution via cron.
Removes IDE friction for non-coding knowledge work; native async/scheduling prevents context pollution and enables parallel agent work. Developers can now run agents on recurring schedules without manual intervention.
Replaces Agent Manager surface from Antigravity IDE with standalone app (macOS/Linux/Windows). Requires fresh install or automatic update from existing IDE. Worth evaluating now if you use agents for tasks beyond code—slash commands (/goal, /schedule, /grill-me) expose core capabilities immediately. IDE remains available during transition.
“Google Antigravity 2.0 is a new, standalone desktop application that fully delivers on a truly agent-optimized experience, available on macOS, Linux, and Windows”
“Dynamic subagents: the main agent can dynamically choose to define and invoke subagents to complete focused subtasks, thereby not polluting the main agents' context window and allowing for parallelism of work”
“Asynchronous task management: Tasks and commands are managed and can run asynchronously to not block the main agent from continuing its work”
“You can now define hooks in a simple JSON format, allowing you to intercept and control the Antigravity agent's behavior”
“millions of developers have adopted the Antigravity IDE”
Get issues like this in your inbox — free, every weekday.
Quick Signals
Nativ runs open models locally on Apple Silicon
MIT-licensed desktop app executes open LLMs directly on Mac hardware via MLX-VLM, eliminating cloud dependencies and API keys.
Developers get real-time telemetry (tokens/sec, memory, thermal state) for local inference tuning, plus endpoint integration for coding agents without vendor lock-in. Eliminates cold-start latency and keeps prompts off external servers.
Replaces cloud-dependent LLM clients (Claude, ChatGPT web) for M-series Macs. Requires macOS + M1+ hardware; supports image/video/audio modalities via MLX-VLM backend. Ready now—open-source, no onboarding friction. Trade-off: model size limited by device RAM.
“Run open models locally on your Mac”
“Built on MLX-VLM and tuned for M-series unified memory and Metal — no wrappers, no translation layers”
“Live tokens/sec, memory pressure, thermal state, and time-to-first-token”
“100% open source”
“No accounts to create, no credits to buy, no data to sell”
Publish an ai-catalog.json on your domain; agents discover and verify tools via federated registries using cryptographic identity tied to domain ownership.
Breaks tool registry silos across organizations—agents can now dynamically locate and trust capabilities at runtime without hardcoded integrations. Replaces fragmented custom registries with a portable, interoperable layer.
Data Point
Apple SpeechAnalyzer beats Whisper Small on accuracy
Apple's new SpeechAnalyzer API ships 2.12% WER on clean speech—outperforming Whisper Small (3.74%) at one-third the compute cost—making it the strongest English on-device option on current Apple hardware.
Developers relying on SFSpeechRecognizer or Whisper for iOS/macOS transcription can now migrate to a faster, more accurate built-in alternative that reduces word error rates by 3.5–4x without external dependencies. The benchmark includes raw transcripts and validation against OpenAI's published Whisper numbers, eliminating guesswork on which engine to ship.
Migrate from SFSpeechRecognizer immediately; it scored 9.02% WER vs. SpeechAnalyzer's 2.12%. Replace Whisper Small for English transcription on Apple platforms (iOS 26+, macOS 26+), but keep Whisper for 100+ language coverage and non-Apple targets. Requires OS version bump and API swap; ready to ship now. Watch for meeting/accented audio benchmarks before assuming parity beyond read speech.
Ready to experiment now: specification is live, quickstart available, Apache 2.0 licensed. Requires hosting ai-catalog.json at a well-known path and supporting registry infrastructure. Production deployment depends on Gemini Agent Platform integration ("coming months"). Worth trying if you operate multi-org agent systems; low friction entry point.
“an open specification for publishing, discovering, and verifying AI capabilities across the web”
“Catalogs: To make resources discoverable, an organization publishes a catalog describing its available capabilities”
“Registries: Registries act as search engines for the agentic web”
“The provider hosts an ai-catalog.json file at a well-known path on its domain”
“ownership of that domain serves as the cryptographic foundation for identity and trust”
Gemma 4 generates tokens 90% faster on Apple Silicon
Ollama's Gemma 4 now uses multi-token prediction with auto-tuning on MLX/llama.cpp backends, delivering ~90% token speedup on macOS without configuration or output changes.
Local LLM inference on Apple Silicon just became practical for coding agents and real-time tasks. Zero-config speedup means existing Ollama workflows get faster automatically—no model retraining or prompt engineering required.
Drop-in improvement for Ollama users running Gemma 4 on M-series chips. Requires updating to Ollama v0.31.1+; no code changes needed. Worth upgrading immediately if you use this model locally.
“generating tokens nearly 90% faster on average across a coding-agent benchmark”
“Ollama auto-tunes how many tokens to draft as it runs, so the speedup is on by default, requires no configuration, and does not change the model's output”
Google AI Edge Gallery adds MCP, scheduled notifications
Mobile agents now call external tools via Model Context Protocol over HTTP, with on-device reasoning and local task scheduling—expanding what Gemma 4 can coordinate without server round-trips.
Developers can wire mobile agents to real data (calendar, email, web) while keeping inference local. Persistent chat history + prompt editing in-app reduce friction for iterating on edge agent workflows.
Replaces manual tool integration patterns; requires MCP server setup (home computer or cloud endpoint). Android ships today, iOS coming soon. Worth prototyping now if you're building local agents—but keep MCP tool descriptions short and data payloads lean due to smaller context windows.
“Google AI Edge Gallery now supports the open-source Model Context Protocol (MCP) over Streamable HTTP as an experimental feature in the Android app”
“By registering a valid MCP URL inside the app, the app dynamically imports tool definitions and resource schemas directly into the on-device model's system prompt”
“prefill speeds can exceed 3,000 tokens per second, allowing the model to restore long session contexts almost instantly”
edge-aimcpmobile-agentsgemmaon-device
Node.js assigns CVE for all EOL versions
Node.js will issue a single CVE identifier for all end-of-life versions, flagged under CWE-1104, to push organizations off unsupported releases via CVE tracking pipelines.
If your organization monitors CVEs for dependency updates, EOL Node.js versions will now trigger alerts in your security tooling, forcing explicit upgrade decisions rather than passive drift. Node.js v16 still gets 11 million downloads monthly despite EOL status.
Run `node -v` and `npx is-my-node-vulnerable` to identify exposure. If you're on v16, v14, or earlier, upgrade to v23, v22, v20, or v18 immediately—this CVE will propagate through standard vulnerability scanners. No workaround; migration is the only option. Act now if you're on EOL.
“Unsupported When Assigned under CWE-1104: Use of Unmaintained Third Party Components”
“Node.js v16, despite being EOL for over a year, has still 11 million downloads per month”