Back to AI Notes

Meta kills Llama for proprietary Muse Spark, Claude Managed Agents go public, Gemma 4 drops

Meta made the biggest strategic pivot in open-source AI history - ditching Llama for a proprietary model called Muse Spark. Anthropic launched Managed Agents in public beta. Google shipped Gemma 4. And OpenAI closed a $122B funding round.

Meta Muse Spark

link

Meta Superintelligence Labs

Meta dropped a bomb on April 8. The company that built its entire AI reputation on open-source Llama models just shipped a proprietary model called Muse Spark. From Meta Superintelligence Labs, the division they created after the $14B deal to bring in Alexandr Wang.

Muse Spark is a natively multimodal reasoning model with visual chain of thought, tool use, and multi-agent orchestration. It's small and fast by design - Meta says it matches their older midsize Llama 4 variant for "an order of magnitude less compute."

The standout capability is health reasoning. They worked with over 1,000 physicians to curate training data. Muse Spark scored 42.8 on HealthBench Hard versus Claude Opus 4.6's 14.8 and Gemini 3.1 Pro's 20.6. Nearly 3x the closest competitor on medical tasks.

It's already powering Meta AI in the Meta AI app, and rolling out to Facebook, Instagram, WhatsApp, Messenger, and the Ray-Ban Meta AI glasses. Meta says they "hope to open-source future versions" but nobody's holding their breath. This feels like the end of the Llama era.

Claude Managed Agents

link

Anthropic

Managed Agents hit public beta on April 8. This is Anthropic's play to own the agent infrastructure layer.

The pitch: a fully managed harness for running Claude as an autonomous agent with secure sandboxing, built-in tools, and server-sent event streaming. Agent loop, tool execution, sandbox container, and state persistence - all bundled into REST APIs. No more building your own agent loops or tool-calling layers.

Claude can read files, run commands, browse the web, and execute code inside the managed environment. You create an agent, configure the container, and run sessions through the API. Server-sent events let you stream the agent's work in real time.

Pricing: standard Claude API token rates plus $0.08 per session-hour for compute. Multi-agent coordination and self-evaluation are still in research preview behind separate access requests.

This is the managed version of what people were building manually with Claude Code and custom harnesses. Anthropic is betting that most developers would rather pay for infrastructure than maintain their own agent runtime.

Google Gemma 4

link

Google

Google shipped Gemma 4 in early April - two models: gemma-4-26b-a4b-it (26B total, 4B active parameters) and gemma-4-31b-it (31B total). Available on AI Studio and through the Gemini API.

Alongside Gemma 4, Google released gemini-embedding-2-preview - their first multimodal embedding model. It handles text, images, video, audio, and PDFs, mapping everything into a unified embedding space. If you're building RAG systems that need to search across different media types, this is significant.

They also dropped Gemini 3.1 Flash-Lite, an efficiency model with 2.5x faster responses and 45% faster output than earlier Gemini versions at $0.25 per million input tokens. Google keeps pushing the "fast and cheap" angle while Anthropic and OpenAI compete on capability.

And a fun one: Google quietly released "Eloquent" - an offline-first AI dictation app on iOS. Real-time transcription, filler word removal, tone adjustment, summarization. All running on-device. Taking on Wispr Flow and SuperWhisper.

OpenAI $122B Round

link

OpenAI

OpenAI closed the largest funding round in history on March 31 - $122 billion from SoftBank, Andreessen Horowitz, Fidelity, and others. Valued at $852 billion.

To put that in perspective: $852B would make OpenAI the 7th most valuable company in the world, ahead of TSMC and behind Amazon. And they raised more in a single round than most countries' annual tech budgets.

The money is going toward compute infrastructure for next-gen models, expanding Codex's capabilities (we saw the results two weeks later), and what they're calling "the next phase of AI" - which based on their recent moves means autonomous agents that can operate computers, manage workflows across days, and work in parallel.

Meta separately announced AI capex of $115-135 billion for 2026 - nearly double the previous year. The capital flowing into AI infrastructure right now is genuinely unprecedented. Whether it's a bubble or the beginning of something transformative depends on which side of the bet you're on.