The Morning Build for November 13, 2026: Apple Foundation Models SDK, Gemini 4 Hydrogen in Search, and $40B Nebius Plan

Today’s brief covers vendor platforms and deployment moves that affect engineering tradeoffs: Apple shipped a native Foundation Models framework and on-device plus Private Cloud Compute paths; Google pushed Gemini 4 Hydrogen into Search and developer surfaces with a 1M-token context and published Orchestrator results; Edge0 open-sourced a 145B on-device model and inference stack; and Nebius disclosed a $40 billion 2027 capex plan with Microsoft capacity delayed into Q2 2027.

Apple releases Foundation Models framework: native Swift API, on-device AFM, Private Cloud Compute, Dynamic Profiles, and macOS fm CLI

What happened: Apple published the Foundation Models framework, a native Swift API that provides access to Apple Foundation Models on-device and in Private Cloud Compute, adds guided generation, tool calling, multimodal prompts, Dynamic Profiles, an Evaluations framework, an fm CLI in macOS 27, a Python SDK, and an Acceptable Use Requirements document.
Why it matters: Engineers building iOS/macOS apps get a single Swift API to target on-device models, server-hosted third-party models, and features for agents and evaluations; Private Cloud Compute access is free for developers enrolled in the App Store Small Business Program with fewer than two million first-time App Store downloads, which changes cost considerations for shipping AI features.
Outlook: macOS 27 availability for the fm CLI is the next concrete milestone that will enable scripting and prototyping without app code, per Apple’s documentation.

Sources: developer.apple.com

Google upgrades Search with Gemini 4 Hydrogen globally, promising more visual, multi-part answers

What happened: Google announced Gemini 4 Hydrogen is available globally in Google Search, delivering improved intent understanding, query fan-out and linking, and more visual, structured responses according to Google’s product post.
Why it matters: Search results will now surface Hydrogen-generated answers across all users, which shifts latency, formatting, and tool-calling behavior in Search outputs engineers rely on for integrations and scraper tooling; the upgrade is live across consumer Search surfaces per Google’s rollout note.
Outlook: The immediate global rollout of Hydrogen in Search, noted as available today by Google, is the product milestone to observe for changes in Search response formats and link surfacing.

Sources: blog.google · the-decoder.com · techcrunch.com

Google publishes Gemini 4 Hydrogen model post: 1M-token context, priced tiers, Orchestrator multi-model claims, and benchmark results

What happened: Google introduced Gemini 4 Hydrogen across developer, enterprise, and consumer surfaces, described four thinking levels, a 1 million-token input and up to 1 million-token output window, pricing at $1.50 per million input tokens and $7.50 per million output tokens, and Orchestrator results showing 24% lower quota and 19% faster completion compared with near-Carbon baselines on Terminal-Bench 4.0.
Why it matters: Developers and enterprises get a high-throughput model option with a 1M-token context window and explicit per-token pricing for large-context workloads; Google also published internal benchmark comparisons and Orchestrator multi-model orchestration metrics that teams can reference when planning multi-model workflows and quota forecasting.
Outlook: Terminal-Bench 4.0 and the Orchestrator metrics Google published are the immediate technical reference points for independent replication and cost-performance comparisons against Hydrogen’s claimed 24% quota reduction and 19% speedup.

Sources: blog.google · antigravity.google · arstechnica.com

Edge0 open-sources Edge0-145B and Edge0-120B plus inference stack, claiming 29GB peak memory and 80+ tokens/s decode on laptops

What happened: Edge0 published Edge0-145B and Edge0-120B with their unified inference framework and deployment tools, reporting a 145B-parameter model that runs with 29GB peak memory, 1,100+ tokens/s prefill, 80+ tokens/s decode, and 128K–256K context support, and released evaluation results comparing multiple benchmarks against named frontier cloud models.
Why it matters: Teams evaluating local-first deployments now have access to published weights, an inference design that pipelines GPU, system memory, and SSD, and internal benchmark claims that place a 145B on-device model in contention with specific cloud models while reducing measured peak memory footprints, which affects decisions about offline agents, data residency, and end-user hardware targets.
Outlook: The technical report and reproduction artifacts Edge0 said it is releasing with the models are the next concrete materials to validate the claimed 29GB peak memory, throughput numbers, and benchmark comparisons.

Sources: edge0.ai

Nebius plans $40 billion 2027 spend funded by prepayments and asset-backed contracts while Microsoft capacity delays push some delivery into Q2 2027

What happened: Nebius disclosed a $40 billion planned spending program for 2027 and said it will rely on customer pre-payments and loans backed by customer contracts to finance construction; the company reported Microsoft capacity at a Vineland, New Jersey site is months behind schedule with just over half planned for delivery by year end and the remainder slated for Q2 2027, per the earnings call and TechCrunch reporting.
Why it matters: Large-scale operator financing tied to prepayments and asset-backed contracts raises timing and availability risk for customers and partners; contract clauses, reported late-delivery penalties, and Microsoft’s ability to terminate affected service after a grace period directly affect capacity guarantees and Nebius’s revenue recognition and cash planning.
Outlook: Q2 2027 delivery of the remaining Microsoft capacity at the Vineland site is the named milestone that will test whether Nebius can meet contract dates and the financing assumptions in its $40 billion 2027 plan.

Sources: techcrunch.com