Nvidia buys Hugging Face for $12.9B

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

🤗 Nvidia buys Hugging Face for $12.9B

🦆 AWS acquires DuckLabs

🔧 AWS adds spec-driven data workflows

⚡ Microsoft releases Agent Lightning v1.0

🤖 Qwen3.8-Flash targets ultra-low cost

🐍 Python 3.15 adds a sampling profiler

Plus: 🎁 5 other news you might like, 🧰 6 tools, and 📚 5 papers.

🤗 Nvidia buys Hugging Face for $12.9B LINK

Nvidia has agreed to acquire Hugging Face, the widely used hub where developers share and download open-source AI models, in a deal that could reshape access to the platform's model hosting and inference tooling.

Owning Hugging Face would hand Nvidia a foothold in the open-source AI ecosystem and a way back into cloud computing, since the platform already lets developers run models on rented compute after Nvidia scaled back its DGX Cloud business.

The talks reportedly hadn't produced a signed agreement and could still fall through, and neither company has commented; the same source says Hugging Face recently reached roughly $150 million in annual revenue, up from about $100 million two months earlier.

🦆 AWS acquires DuckLabs LINK

AWS is acquiring DuckLabs, the Amsterdam team behind DuckDB, the embeddable open-source analytical database often called the SQLite of analytics, with the deal expected to close in September 2026.

AWS committed to keeping DuckDB and ecosystem projects like DuckLake and Quack open-source under MIT licenses, with the core team staying in Amsterdam under co-founders Hannes Mühleisen and Mark Raasveldt and the nonprofit DuckDB Foundation as steward.

The collaboration builds on 2025 work adding DuckDB support to Amazon S3 Tables and SageMaker Lakehouse, and AWS plans a technical advisory board within the Foundation plus signed third-party extension support for the ecosystem.

🔧 AWS adds spec-driven data workflows LINK

AWS has published a specification-driven composition pattern for data workflows that separates workflow intent from processing logic, cutting duplicated pipeline code and easing validation and governance across growing sets of datasets and variants.

Specifications written in JSON or YAML describe source and target datasets, field mappings, and transformations, while a Lambda composer validates them, queries OpenSearch for capability metadata, and generates a Step Functions state machine invoking reusable processors.

A capability registry holds versioned, discoverable transformation functions, and specs can tag sensitive fields so the composer validates data classification and emits masking artifacts, though AWS warns the pattern adds needless complexity for simple or low-volume workflows.

⚡ Microsoft releases Agent Lightning v1.0 LINK

Microsoft shipped Agent Lightning v1.0 on GitHub, an MIT-licensed framework that lets teams run reinforcement learning on production agents without reimplementing the agent loop inside a separate training environment.

The approach hands the production harness ownership of context construction, tool execution, and the agent-environment loop, while the trainer only observes request-response pairs across a service boundary, cutting train-serve mismatch that drifts behavior in production.

Using 6K training examples, it lifted Qwen3.5-9B on SWE-bench Verified from ~42% to ~56%, and the whole framework is roughly 3,500 lines of core Python plus data-cleaning and reward-hacking prevention, though it realistically needs GPU and Kubernetes clusters.

🤖 Qwen3.8-Flash targets ultra-low cost LINK

Alibaba released Qwen3.8-Flash, an open-weight multimodal MoE model and early preview of the Qwen4 architecture, aimed at ultra-low cost pricing set at 16 cents per million input tokens and 47 cents per million output tokens via the QwenCloud API.

The model carries 125 billion parameters plus 51 billion N-gram embeddings, activating just 6 billion per token, and was trained for a ninth of the cost of Qwen3.7-Plus while beating it on coding and office tasks.

The production version ships soon through the QwenCloud API, extending a Qwen lineup that has open-sourced over 460 models and passed 3 billion downloads in six months, ahead of Meta, Google, and DeepSeek.

🐍 Python 3.15 adds a sampling profiler LINK

Python 3.15 introduces a new profiling package under PEP 799, adding a statistical sampling profiler code-named Tachyon that pinpoints where time goes with virtually no runtime overhead, even against live processes.

The sampler reads a program's call stack from the outside, letting you attach to a running process by PID without restarting or modifying it, given permissions, with modes separating CPU work, I/O waits, and GIL-held code.

Output formats include interactive flame graphs, line-level heatmaps, and a live top-style terminal dashboard, while the reorg keeps cProfile as an alias under profiling.tracing and deprecates the old profile module for removal in Python 3.17.

Other news you might like

• Haiku R1/beta6 releasedLINK

• Inside SWSH's automated mobile release pipeline with ExpoLINK

• Raspberry Pi now has an official guide for building your own cyberdeckLINK

• Patches Provide For Much Faster In-Kernel Zstd Due To Embarrassingly Bad InefficiencyLINK

• Armbian 26.8 releasedLINK

🧰 Trending tools

AnySearch: queries multiple trusted sources in parallel for AI agents, returning filtered, de-duplicated results as structured data ready to useLINK

SKI: adds voice control to AI coding agents like Claude Code and Codex, letting you dictate commands and hear responses hands-free while coding.LINK

Dograh: open-source voice agent platform with a visual flow builder, telephony, human transfer, and self-hosted QA monitoring across 30+ model integrations.LINK

Zro: routes coding requests to open-source models like GLM-5.2, DeepSeek V4 Flash, and Kimi K3 through one endpoint without storing your dataLINK

Meridian: open-source activity tracker that runs locally on your device, logging work with context so you can recall where time actually went.LINK

AgentSky: runs managed AI coding agents with auto-recovery, controllable from WhatsApp, Telegram, Slack, web, API, or CLI for remote development workflows.LINK

📚 Trending papers & reports

Coding-agent supervision grades software fixes by comparing candidate patches against each other instead of running fragile tests, boosting an open model's first-try fix rate by ~10 points, up to ~33% overall.LINK

Science task benchmark shows the best AI agents fully finished only about 21% of 97 real research workflows, and three-quarters of failed attempts still falsely claimed completion, exposing a reliability gap for automated science.LINK

Website design specs turns screenshots of public sites into auditable, one-per-source design documents that flag what's measured versus guessed and never mix evidence between references, giving teams cleaner blueprints from visual examples.LINK

Spec-driven coding data gives the first large-scale look at how developers use AI tools that turn written specifications into code, cataloging over 470,000 spec files across 73,000 public projects to reveal how specs become working software.LINK

AI agent safety layer checks every action a tool-using agent tries before it can change files or systems, blocking unauthorized moves while cutting median response time from ~26 to ~14 milliseconds.LINK

See you tomorrow for a new dose of ☕️ Devshot!