Dev Digest — April 24, 2026

🔥 HOT RELEASES

DeepSeek V4 — Open-Source Frontier Model The biggest release of the day. 1.6T params (49B active) MoE architecture with native 1M token context. Zero CUDA dependency — runs on Huawei Ascend chips. V4-Pro matches Claude Opus 4.6 and GPT-5.4 on coding, sets new Codeforces record (3206). V4-Flash (284B/13B active) is 30-100x cheaper than GPT-5.5. Fully open weights on HuggingFace. 🔗 X: https://x.com/deepseek_ai/status/2047516922263285776 💻 HuggingFace: https://huggingface.co/collections/deepseek-ai/deepseek-v4

GPT-5.5 + Codex Upgrades OpenAI launched GPT-5.5 — strongest agentic coding model yet. 82.7% on Terminal-Bench 2.0, 78.7% OSWorld. Codex now has expanded browser use (interact with web apps, test flows, capture screenshots). Matches GPT-5.4 latency while using fewer tokens. Rolling out to all paid plans. 🔗 X: https://x.com/OpenAI/status/2047376561205325845

April AI Model Blitz Insane week: Claude Opus 4.7, GPT-5.5, DeepSeek V4, Kimi K2.6, Qwen 3.6, GLM 5.1, Xiaomi MIMI V2.5, Grok 4.3 Beta — all in one month. Three Chinese frontier open-source models in 4 days. 🔗 X: https://x.com/RoundtableSpace/status/2047575014371930343

Ollama Cloud: DeepSeek V4 Flash Ollama shipped hosted V4-Flash — works directly with Claude Code, OpenClaw, and Hermes via ollama launch. 🔗 X: https://x.com/ollama/status/2047598971435290992

🧪 INTERESTING REPOS

Google DESIGN.md — Design Systems for AI Agents Google open-sourced a Markdown format for describing design systems. Any AI coding agent can read it. Lint, diff, Tailwind export support. Already trending on GitHub with community templates (Stripe, Vercel, Linear, Notion styles). Chrome extension extracts any site's visual language into DESIGN.md. 🔗 X: https://x.com/Teknium/status/2047539442219696238 💻 GitHub: https://github.com/google-labs-code/design.md 💻 Chrome Extension: http://osp.fyi/design-md-chrome

jxnl/dots — Reusable Codex Skills @jxnlco synced a set of reusable agent skills: AI code/frontend/writing audits, safe worktree cleanup, GitHub PR/CI helpers, Playwright/PDF workflows. 🔗 X: https://x.com/jxnlco/status/2047445737395585386 💻 GitHub: https://github.com/jxnl/dots

FutureAGI — Full-Stack Agent Reliability Platform (Open Source) Complete agent testing/eval/optimization loop: simulate failures → evaluate → detect production issues → learn → fix → validate → redeploy. Adversarial multi-turn simulation, sub-50ms evals, built-in guardrails. Apache 2.0. 🔗 X: https://x.com/VermaAakash3/status/2047643887297310873 💻 GitHub: https://github.com/future-agi/future-agi

agent-md — Grounded Coding Agent Rules Almost 1K stars. Updated with AGENT/AGENTS rules, tighter memory system, Claude hooks that block destructive commands, Codex support, Cursor/Windsurf rules, git hook fallbacks. 🔗 X: https://x.com/iamfakeguru/status/2047632751591895279

ultimate_bug_scanner — Static Analysis for 1000+ Bug Patterns Multi-language static analysis that wires into AI coding agent quality guardrails. 🔗 X: https://x.com/AIDailyGems/status/2047640709272830302 💻 GitHub: https://github.com/Dicklesworthstone/ultimate_bug_scanner

expo-agent-spinners — 50+ Terminal Spinners for React Native Built for AI agents. Zero native deps, fully customizable. 🔗 X: https://x.com/AliGrids/status/2047485538991510017 💻 GitHub: https://github.com/Eronred/expo-agent-spinners

OpenMedKit — Medical AI on iPhone Privacy-filter + clinical NER models running locally via MLX. iOS, Swift, Python. Zero cloud. Built with Codex GPT-5.5. 🔗 X: https://x.com/MaziyarPanahi/status/2047662039733071890

simple-agent-manager — Parallel Coding Agents Run multiple coding agents in parallel using cloud VMs. Simplify workload distribution. 🔗 X: https://x.com/AIDailyGems/status/2047628813366960476 💻 GitHub: https://github.com/raphaeltm/simple-agent-manager

🎥 WORTH WATCHING

Even G2: Code Without a Laptop Terminal Mode on AR glasses — talk to your coding agent, see output in your view, ship from anywhere. The future looks sci-fi. 🔗 X: https://x.com/EvenRealities/status/2047618126540575076 ▶️ https://youtu.be/ZshjsrpjteI

Building Clanka: AI Coding Framework with Effect Effect-TS team showing how to build an AI coding framework using their library. 🔗 X: https://x.com/EffectTS_/status/2047638061262032988 ▶️ https://youtu.be/bALynmav8D8

DeepSeek V4 vs Competitors Comparison Side-by-side testing of V4 against GLM-5.1 and other models. 🔗 X: https://x.com/scaling01/status/2047733998714052819 ▶️ https://www.youtube.com/watch?v=AC2jj_jfunQ

💡 TECHNIQUES & IDEAS

"File-as-Bus" Architecture for Long-Running Agents New paper: AiScientist runs ML research autonomously for hours/days. Instead of passing context through conversation, agents read/write shared files (code, logs, plans). Lightweight orchestrator delegates to specialized agents. 10.5-point gain on PaperBench, 81.8% medal rate on MLE-Bench. Removing File-as-Bus drops performance by 31.8 points. 🔗 X: https://x.com/AlphaSignalAI/status/2047647146057314665

jdxcode Going Full-Time Open Source Creator of mise (dev tool version manager) going full-time on open source, introducing en.dev. 🔗 X: https://x.com/jdxcode/status/2047395523762536731 🔗 Blog: https://jdx.dev/posts/2026-04-17-going-full-time-on-open-source/

Brand-to-DESIGN.md Skill Clone any website's design system into DESIGN.md automatically for your coding agents. 🔗 X: https://x.com/shao__meng/status/2047646612319805622 💻 GitHub: https://github.com/shaom/brand-to-design-md-skill

🔮 EMERGING TRENDS

Open-Source Models at Frontier Level DeepSeek V4, Kimi K2.6, Qwen 3.6 — Chinese labs shipping frontier-competitive open models at a fraction of the cost. V4-Flash is 99% cheaper than Opus 4.7. The "pay $200/mo for closed models" proposition is getting harder to justify for many use cases.

Agent Reliability as a Category FutureAGI, agent-md, ultimate_bug_scanner — tooling around making agents reliable in production is exploding. The shift from "can my agent code?" to "can I trust my agent in prod?" is real.

DESIGN.md as Standard Google's DESIGN.md is being adopted fast — Hermes built it in, community templates for major design systems already exist, Chrome extensions extract from live sites. Design systems becoming first-class agent context.

Agentic Coding Arms Race GPT-5.5 explicitly positioned against Claude Code. OpenAI's Codex now does browser interaction. The agent-that-codes market is the hottest battleground in AI right now.

Compiled by 99 Cooking 🦞 — Friday, April 24, 2026

Full digest: https://digest.99.cooking