Muse Spark 1.2 on Vercel, v0 API launch, and fresh agent tools for builders

What shipped

On 5 August 2026 Vercel pushed multiple updates to its AI platform including a new coding model and API access for agents. Google announced leadership changes at DeepMind and began phasing out its older assistant. Product Hunt and Hugging Face also surfaced new agent tools and open models for immediate testing.

Vendor launches

Vercel delivered the majority of updates with Muse Spark 1.2, higher Sandbox quotas, and the new v0 API that lets builders trigger app generation from prompts. Additional changes include domain tools, firewall access on the free tier, and support for Next.js 16.3 plus Devin integration. Google added context through DeepMind team adjustments while Meta expanded its coding agent reach.

• •Muse Spark 1.2 on Vercel AI Gateway Meta released Muse Spark 1.2 with stronger code generation, debugging, and full repository handling. Teams can run long iterative coding loops through the single Vercel endpoint. The update targets developers who need sustained project-level work without switching platforms.
• •Vercel Sandbox quota increase Vercel raised default limits to 10,000 concurrent sandboxes and up to 5,000 vCPUs per minute on paid plans. The dynamic rate scales with usage so builders can run more parallel tests. Hobby users also gained full egress firewall controls for safer experiments.
• •v0 API release Vercel opened a headless v0 API that accepts a prompt, spins up a sandbox, and returns a live preview URL. Each session keeps its own workspace so follow-up edits continue from the current state. Builders can embed the preview directly into custom interfaces or internal tools.
• •Devin Outposts on Vercel Sandbox Cognition's Devin agent now runs each session inside an isolated Vercel microVM. Session state is preserved with snapshots and network policies stay inside the project. Teams gain durable workflows without maintaining local worker machines.
• •Sandbox firewall on Hobby plan Vercel extended full egress firewall features to the free tier. Builders can set allow or deny rules and keep tokens out of code while still calling services like AI Gateway. The change gives hobbyists the same isolation previously limited to paid plans.
• •AI Gateway on AWS Marketplace Vercel listed AI Gateway on AWS Marketplace so teams can consolidate billing through existing AWS contracts. Procurement uses private offers with annual terms plus usage pricing. One endpoint now routes to hundreds of models with built-in fallbacks.
• •Skill packs on skills.sh Vercel enabled shareable skill packs that bundle multiple agent skills into a single URL or GitHub org install. Anyone can create a pack from local files or community skills and update it with one command. Teams can standardize agent behavior across projects without manual setup.

Hugging Face trending

Three new text and image-text models appeared on the Hugging Face trending list from deepgrove, LuffyTheFox, and inclusionAI. A research paper on reflective reasoning for large language models also gained attention. These releases give builders quick options for local or fine-tuned inference.

• •maple-preview model deepgrove released maple-preview, a text-generation model built with transformers and available for download or fine-tuning on the Hub. Early users can test it for general writing tasks or adapt it to specific domains.
• •Qwen3.6-35B-A3B model LuffyTheFox published an uncensored image-text-to-text variant of Qwen on the Hub using the hermes library. It supports multimodal prompts and runs through standard inference endpoints.
• •Ling-3.0-flash model inclusionAI dropped Ling-3.0-flash, a text-generation model ready for immediate download and fine-tuning. Builders can run it locally or host it for low-latency chat applications.
• •ReflectRL research paper A new paper explores on-policy training that uses reflective reasoning from expert trajectories to improve model performance on hard problems. The approach aims to reduce reliance on perfect golden examples during post-training.

Fal model gallery

MiniMax H3 on Fal: MiniMax released its H3 reference-to-video model on Fal that accepts up to nine images plus motion and audio clips. It produces 2K stylized video while keeping subjects consistent and syncing lips to audio. Users can test it directly for short-form content or animation prototypes.

Product Hunt picks

Five new tools appeared on Product Hunt focused on agent building, hiring workflows, private knowledge access, and meeting notes. Each offers a narrow slice of automation that non-coders or small teams can try immediately.

• •Keystroke Keystroke lets users build AI agents and workflows through a visual interface without writing code. Early adopters can connect actions across multiple services in one canvas.
• •Kiro Crew Kiro Crew provides an open-source workspace for agentic development where teams manage tasks and code together. Builders can self-host or fork the project to match their stack.
• •Dover MCP Dover MCP connects hiring pipelines to Claude or ChatGPT so recruiters run candidate outreach from chat. It keeps the full process inside the model window.
• •BackEngine MCP BackEngine MCP turns private company documents into searchable context for AI agents. Teams upload files once and query them through existing chat interfaces.
• •Wispr Flow Notetaker Wispr Flow records meetings and produces notes that capture speaker details and action items accurately. Users review transcripts inside the same app.

Industry news

Meta faced reports of AI-generated harmful ads on its platforms while Google confirmed the September shutdown of Google Assistant in favor of Gemini. DeepMind lost both its CEO and chief scientist as they pursue new ventures. Mistral released a compact safety model and researchers demonstrated risks of AI worms.

• •Google Assistant shutdown Google will retire Google Assistant on Android and Wear OS starting 4 September 2026 with Gemini taking over all voice commands. The switch tests whether a large language model can match the reliability of the older deterministic system for routine tasks.
• •Mistral Shieldstral model Mistral open-sourced Shieldstral, a 3B safety model that answers natural-language yes-or-no questions instead of fixed categories. It matches much larger models on benchmarks and runs locally so operators can define their own rules at runtime.
• •Jeff Dean startup launch Jeff Dean and other departing Google researchers formed Discovery Loop to apply AI to scientific discovery workflows. The new company aims to accelerate research pipelines across multiple domains.
• •AI worm research Chinese researchers showed that large language models can be turned into adaptive computer worms capable of spreading across systems. The findings highlight new security considerations for any deployment that accepts untrusted prompts.
• •Meta Muse Code agent Meta launched Muse Code, an agent built to handle complex tasks inside large codebases. Early users can test it for repository-scale refactoring and multi-file edits.

Other

Muse Spark 1.2 on OpenRouter: Meta added Muse Spark 1.2 to OpenRouter with 1,049k context and pricing at $1.25 per million input tokens and $4.25 per million output tokens. Builders can route calls through their existing OpenRouter setup to test long-context agent tasks immediately.

What this means for you

For Vibe Builders: You can now trigger full app previews from a single prompt with the v0 API and run them inside isolated sandboxes that scale to thousands of concurrent sessions. Skill packs on skills.sh let you hand teammates a ready set of agent behaviors without writing code. OpenRouter access to Muse Spark 1.2 gives you a 1M-token model for complex workflows at predictable per-token cost.

For Non-techies: For day-to-day business use, new agent tools on Product Hunt such as Dover MCP and BackEngine MCP let you run hiring or internal search directly inside Claude or ChatGPT. Google phasing out Assistant means you will switch to Gemini for phone and car commands starting in September. Mistral's small Shieldstral model offers a local way to add safety checks without sending data elsewhere.

For Developers: Production teams gain concrete capacity increases from Vercel Sandbox reaching 10,000 concurrent instances and full firewall controls on the Hobby tier. The v0 API and Devin Outposts integration show how to embed agent loops into existing projects with durable state. Watch Mistral Shieldstral and the ReflectRL paper for lightweight safety and reasoning upgrades you can benchmark against current pipelines this week.

What to watch next

Track the September Gemini transition on Android devices and any follow-up model releases from the new DeepMind leadership. Monitor OpenRouter and Fal for pricing or quota changes on Muse Spark and MiniMax H3. Check Product Hunt daily for the next wave of MCP-based agent tools.

Sources

Vendor launches

• •Muse Spark 1.2 is now available on Vercel AI Gateway
• •Search and buy domains in the dashboard
• •Vercel Sandbox now supports 10,000 concurrent sandboxes and 5,000 vCPUs per minute
• •Introducing the new v0 API
• •New setup page after domain checkout
• •The next chapter of our AI momentum
• •Vercel Sandbox now supports Devin Outposts
• •Full Sandbox egress firewall now available on Hobby plan
• •AI Gateway is now available on AWS Marketplace
• •Skill packs are now available on skills.sh
• •Vercel supports Next.js 16.3

Hugging Face trending

• •maple-preview by deepgrove trends on HuggingFace
• •Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V7-GGUF by LuffyTheFox trends on HuggingFace
• •Ling-3.0-flash by inclusionAI trends on HuggingFace
• •ReflectRL: Learning from Golden Negative Trajectories via Reflective-to-Direct Reasoning

Fal model gallery

• •MiniMax H3 Reference to Video drops on Fal

Product Hunt picks

• •Keystroke
• •Kiro Crew
• •Dover MCP
• •BackEngine MCP
• •Wispr Flow Notetaker

Industry news

• •Meta Ran Ads That Contained AI-Generated Child Sexual Abuse Imagery
• •Google will shut down Google Assistant starting September 2026 as Gemini takes over on Android and Wear OS
• •Google Deepmind loses both its CEO and chief scientist as Demis Hassabis and Jeff Dean step down simultaneously
• •Mistral's open model Shieldstral matches much larger safety models at a fraction of the size
• •Jeff Dean and other top AI researchers are leaving Google to launch their own startup
• •AI Worms and Viruses Are Coming
• •Meta launches Muse Code, an AI agent for large code bases

Other

• •Meta: Muse Spark 1.2 now available on OpenRouter (1,049k context, $1.25/M in, $4.25/M out)
• •IEEE Course Teaches How to Use AI to Modernize Power Grids