☕️ DeepSeek V4-Flash is world's cheapest model

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

🐋 DeepSeek V4-Flash is world's cheapest model

⚡ TypeScript 7.0 ships a native Go compiler

🤖 Microsoft Agent Framework hits general availability

🚢 Kubernetes v1.37 sneak peek is here

🔑 Docker adds OIDC for GitHub Actions

💾 New tool case-folds source code at memory speed

Plus: 🎁 13 other news you might like, 🧰 6 tools, and 📚 5 papers.

🐋 DeepSeek V4-Flash is world's cheapest model LINK

DeepSeek released V4-Flash as open weights on Hugging Face, a 304B-parameter model with a 167GB footprint that you can run through OpenRouter or the DeepSeek API, priced at $0.14 per million input tokens.

DeepSeek says this release has substantially enhanced agentic capabilities over earlier V4 models, and Artificial Analysis ranks it ahead of MiniMax M3, a larger 428B model, on the Intelligence Index.

One practical inference note: the default reasoning level produced weak output during testing via OpenRouter, so set reasoning to high for usable code generation, though that raises token cost by roughly six times.

⚡ TypeScript 7.0 ships a native Go compiler LINK

Microsoft has released TypeScript 7.0, the first stable build to ship a native compiler ported to Go, delivering full-build speedups typically between 8x and 12x while preserving complete type-checking.

On the VS Code source, a full build dropped from 126 seconds to 10.6 seconds with ~18% less memory, and Slack cut CI type-checking from 7.5 minutes to 1.25 minutes.

The release ships without a stable programmatic API, expected in 7.1, so typescript-eslint, webpack loaders and framework tooling for Vue, Svelte, Astro and Angular can't use it yet.

🤖 Microsoft Agent Framework hits general availability LINK

Microsoft's Agent Framework reached 1.0 general availability on April 2, 2026, shifting from an SDK into a supported production runtime that runs as a single binary across local development, containers, and hosted deployment in.NET and Python.

The Agent Harness bundles function invocation, per-call history, context compaction, plan-and-execute todo lists, file memory, web search, tool approval, and OpenTelemetry by default, while shell access, background sub-agents, and automatic looping stay opt-in and emit warnings.

New GitHub Copilot SDK and Claude Agent SDK connectors let orchestrations delegate to third-party coding agents without custom adapters, routing their traffic through the same identity, content safety, and observability policies already governing the fleet.

🚢 Kubernetes v1.37 sneak peek is here LINK

Kubernetes v1.37, due August 26th, 2026, has a sneak peek out detailing planned deprecations, breaking changes, and graduations engineers should prepare for across kubectl, kubelet, kube-proxy, and storage health.

kube-proxy's ipvs mode is being deprecated and now logs a startup warning, with default disablement expected by v1.40 and full removal by v1.43, while cgroup v1 support is also on the removal path after failCgroupV1 defaulted to true.

The metrics.k8s.io API graduates to GA after nearly nine years in Beta with no functional changes, kubelet in User Namespace rootless mode reaches Beta, and SELinuxMount hits GA, potentially breaking pods sharing volumes with differing SELinux labels.

🔑 Docker adds OIDC for GitHub Actions LINK

Docker now supports OpenID Connect for GitHub Actions, letting workflows authenticate to Docker Hub with short-lived, per-run tokens instead of stored personal or organization access tokens that require manual rotation and risk leaking.

GitHub issues a signed JWT that `docker/login-action` presents to Docker, which verifies it against rulesets in the Admin Console and returns a scoped token that expires in minutes and cannot be reused.

Setup requires an OIDC connection in Docker Home with up to five rulesets pinning repos and branches, plus adding `id-token: write` permission and a connection ID to your workflow YAML; existing tokens keep working during migration.

💾 New tool case-folds source code at memory speed LINK

GitHub's Blackbird code search engine now case-folds text at over 45 GiB/s using a new open-source Rust crate called `casefold`, indexing 180 million repositories and more than 480TB of source code.

The key ASCII speedup came from deleting an early-exit `break`, since a data-dependent loop exit blocks vectorization; a fully branchless sweep lets LLVM emit 16-byte NEON writes, jumping from 3.1 GiB/s to memory-bandwidth speed.

The crate implements only simple 1-to-1 folds (matching ripgrep), avoids heap allocation by mutating the input `String` in place, and packs Unicode 16.0's 1484 fold mappings into 1776 bytes with a page bitmap for cheap negative tests.

Other news you might like

• I'm (mostly) picking models on speed now, not intelligenceLINK

• Stop graphing everything: When GraphRAG actually beats vector RAGLINK

• Institutional knowledge doesn't scale: Building an agentic data analystLINK

• How fast is C++26’s std::hive?LINK

• The Conductor DeveloperLINK

• HashiCorp Ships Public Beta of Vault Kubernetes Key ManagementLINK

• Cloudflare Introduces Meerkat for Strongly Consistent Global CoordinationLINK

• First open-source firmware for AM5 officially launches, Dasharo v0.9.0 brings Coreboot and openSIL to Zen 4 APUs on MSI B850LINK

• Microsoft Paint used as a 'monitor' to run Doom at up to 35 fps, project released by firm's Azure CTO, runs actual Doom engine and loads real shareware DOOM1.WADLINK

• How to build an MCP app on the 2026-07-28 spec with WorkOS AuthKitLINK

• Linux 7.2-rc6 was a massive release, but it's not going to delay the final releaseLINK

• NetBSD 11.0 is here with RISC-V support and Snapdragon X Elite improvementsLINK

🧰 Trending tools

AnySearch: a search API for AI agents that pulls filtered, de-duplicated, structured results from trusted sources in parallel, improving reliability.LINK

Sim: a workspace for building and deploying AI agents visually or with code, connecting to 1,000+ integrations and every major LLM provider.LINK

Prelint: reviews AI-generated pull requests against your ADRs, docs, and past decisions to catch product drift before code ships.LINK

Zro: routes coding requests to open-source models like MiniMax M3, GLM-5.2, and Kimi K2.7 across regions without retaining any data.LINK

FetchSandbox: simulates webhooks, retries, and async workflows across 60+ real APIs, letting developers reproduce bugs deterministically without burning API quota or touching staging.LINK

Cekura: automates QA testing for conversational AI agents, covering pre-production simulation, evaluation, and production call monitoring with CI/CD pipeline integration.LINK

📚 Trending papers & reports

Automated test-writing for AR apps uses AI to generate and refine the pass/fail rules testers need, resolving conflicting suggestions in 88.2% of cases and producing checks valid enough to catch real coding errors.LINK

Digital twin data contracts let companies set explicit quality rules, like accuracy and timeliness, for the data feeding their simulations, so bad inputs get caught automatically instead of silently corrupting predictions.LINK

Coding assistants that jump to edit code before gathering enough evidence get a built-in checklist forcing them to verify context first, boosting task success by up to ~12 points while cutting token costs up to ~12%.LINK

AI-built process-mining software was coded almost entirely by an AI coding agent over 65 hours and 151 commits, yet fixing the agent's own mistakes outnumbered adding new features by more than 2 to 1.LINK

C-to-Rust migration uses AI agents to rewrite legacy C code, first into safe but clunky Rust, then into clean idiomatic Rust, testing each step, successfully converting a 12,500-line real-world program.LINK

See you tomorrow for a new dose of ☕️ Devshot!