☕️ DeepSeek launches V4 Pro

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

🐋 DeepSeek launches V4 Pro

🎵 Spotify speeds up its data lake queries

🐳 Docker's new VMM enters public beta

⚡ Astro 7 builds up to 61% faster

🌐 Databricks configures tens of millions of VMs

🐍 Simon Willison releases alchemy-utils 0.1

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

🐋 DeepSeek launches V4 Pro LINK

DeepSeek quietly finalized its flagship V4 Pro, shipping the DeepSeek-V4-Pro-0813 build with no blog post-the only signal being an updated version string on its API pricing page after months of the model running as a preview.

Pricing holds steady at roughly $0.44 per million input tokens and $0.87 per million output, but the weights underneath changed; the earlier scores every independent lab tested described the unfinished preview, so nobody has externally benchmarked 0813 yet.

DeepSeek's own numbers across 10 agent benchmarks trail Fable 5 by about 5% on average, but on blended rates V4 Pro costs around 46 times less; the MIT-licensed weights sit on Hugging Face for independent verification.

🎵 Spotify speeds up its data lake queries LINK

Spotify built Random Access Parquet (RAP), a storage layer that serves low-latency point queries directly against its data lake, letting online services and AI apps fetch individual records without replicating datasets into operational databases like Bigtable.

RAP adds an external index mapping keys such as user IDs to specific Parquet files and row locations, so a lookup resolves through the index and issues a targeted ranged read instead of scanning thousands of files.

Storage optimizations-sorting by lookup key, interleaving columns, and covering indexes-trade modest size increases for fewer operations, serving some point queries in a single few-kilobyte read, while secondary indexes enable multi-dimension queries without rewriting data.

🐳 Docker's new VMM enters public beta LINK

Docker has launched a public beta of Docker VMM, a first-party virtualization layer it built from scratch to replace the third-party VMM under Docker Desktop, available on Mac and Windows in v4.86.

Developers get faster container startup, quicker host-container file sharing, and memory that returns to the host when containers idle, while Windows gains Hyper-V-level isolation with WSL2-like speed for the first time under a Docker-maintained VMM.

Mac users already on Docker VMM upgrade automatically; Windows users opt in via Settings > General with no waitlist, Linux support arrives at GA, targeted for late October 2026 when Docker VMM becomes the default for new installs.

⚡ Astro 7 builds up to 61% faster LINK

Astro 7 shipped with its build pipeline rewritten in native code, cutting build times by 15 to 61 percent, with the astro.build site dropping from 63s to 24s and Cloudflare's 8,400-page docs falling from 387s to 262s.

The compiler moves from Go to a Rust build on oxc and Lightning CSS, Markdown now runs through a Rust processor called Sätteri, and the framework sits on Vite 8 with Rolldown, though remark and unified pipelines remain available.

Compilation turned stricter, erroring on unclosed tags and passing invalid nesting through as written, while point releases added CSP directives for inline styles, per-route incremental builds in 7.2, and JSON logging for AI coding agents.

🌐 Databricks configures tens of millions of VMs LINK

Databricks re-architected how it delivers network configuration to the tens of millions of VMs its serverless platform launches daily, cutting p99 RPC latency by 97.5% from 5,000ms to 125ms while hitting 99.99% availability.

The old design synchronously fetched config from multiple upstream services on the cluster-start critical path; the rewrite subscribes to change events via a message queue and pre-computes per-workspace snapshots, so serving becomes a single storage read.

A periodic reconciler re-syncs all workspaces as a safety net for missed events, and calling upstream only on actual changes rather than every request dropped upstream call volume by 86% across billions of daily requests.

🐍 Simon Willison releases alchemy-utils 0.1 LINK

Simon Willison has shipped alchemy-utils 0.1a0, an early alpha of a database-agnostic library and CLI that ports his sqlite-utils API to SQLAlchemy so it works across PostgreSQL, SQLite, and DuckDB.

The tool mirrors sqlite-utils core methods like insert, upsert, insert_all, upsert_all, create, update, and table introspection, and runs via uvx with per-engine extras such as alchemy-utils[postgresql] or alchemy-utils[duckdb].

Willison built the prototype using Codex and GPT-5.6 with red/green TDD and pytest; a DuckDB CSV insert of every San Francisco tree initially took nearly an hour, but Codex optimized it down to about 35 seconds.

Other news you might like

• The web’s newest weapon against AI scrapers is a fontLINK

• How We’re Building Scam Alert on WhatsApp With End-to-End Encryption and Verifiability GuaranteesLINK

• Proton VPN is replacing wireguard-go with a new Rust VPN coreLINK

• Netflix Adopts Cloud-Native Job Queueing System Kueue to Replace an In-House SolutionLINK

• Okta targets AI agent token costs with MCP scopingLINK

🧰 Trending tools

AnySearch: a search API for AI agents that queries multiple trusted sources in parallel, then filters and de-duplicates results into structured data.LINK

Prefactor: an evaluation layer that scores agent runs in real time, catching quality regressions and drift before they impact customers in production.LINK

ZooData: converts webpages into structured JSON, cutting LLM token usage by 75% while delivering pre-analyzed Amazon and TikTok e-commerce data for AI agents.LINK

CartAI: automates online checkout across merchant sites through a single API call, handling payments, shipping, and order tracking without per-site integrations.LINK

ditto.site: turns any public URL into clean Next.js or Vite code deterministically, preserving components, design tokens, and hover states via a free API or MCP server.LINK

Replay QA: records website sessions with devtools access, letting you create detailed, shareable bug reports without running manual QA processes.LINK

📚 Trending papers & reports

Automated bug-hunting for Python software libraries reads their documentation with an AI model to generate valid test inputs, catching 74 issues, with 43 confirmed and 29 already fixed by developers.LINK

Multi-agent coding tools were tested head-to-head on a real task, summarizing README files, and showed no meaningful performance difference between frameworks, though most still lack basic monitoring features developers need.LINK

Code merge conflicts get resolved more accurately across programming languages by a tool that reads code structure only where needed, cutting false alarms without slowing builds or missing real conflicts.LINK

Coding agent toolkits show that how tools are packaged, not just what they can do, matters, since better-structured interfaces made repeated attempts up to 4.7x more consistent, cut steps by ~42%, and lowered token use by ~56%.LINK

Code-writing AI trained with formal proof reasoning behind correct solutions solves far more hard programming tasks, boosting correct answers from 16 to 23 on a tough benchmark, a 44% jump over prior methods.LINK

See you tomorrow for a new dose of ☕️ Devshot!