Dev Digest — April 21, 2026
🔥 HOT RELEASES
Kimi K2.6 — Moonshot AI's 1T-parameter open-weight model The big release of the day. 1 trillion parameters (32B active via MoE with 384 experts), Apache 2.0 licensed. Claims SWE-bench Pro 58.6% (beating GPT-5.4 and Claude Opus 4.6). The headline feature is Agent Swarm — up to 300 sub-agents running 4,000+ coordinated tool calls. Already available in Notion and Warp. During development, the model deployed Qwen3.5-0.8B locally, coded its own Zig inference engine, and improved throughput from ~15 to ~193 tokens/sec — faster than LM Studio. 🔗 X: https://x.com/akothari/status/2046423772048314480 🔗 X: https://x.com/warpdotdev/status/2046629258768925143 🔗 X: https://x.com/Hesamation/status/2046269399900860922
Google Stitch — design.md: Open-source visual language for AI design Google Labs open-sourced a spec + CLI for a shared design language aimed at AI-driven UI workflows. Called "design.md", it standardizes how design components are described so AI tools can generate consistent UIs. 🔗 X: https://x.com/stitchbygoogle/status/2046624733828108570 💻 GitHub: https://github.com/google-labs-code/design.md 📝 Blog: https://blog.google/innovation-and-ai/models-and-research/google-labs/stitch-design-md/
Gemma 4 — 10 concurrent instances on a MacBook Google open-sourced a demo showing Gemma 4 26B A4B running 10+ concurrent requests on M4 Max at 18 tok/s per request. Impressive efficiency for local multi-agent setups. 🔗 X: https://x.com/googlegemma/status/2046621841146671456
🧪 INTERESTING REPOS
OpenCut — Privacy-first open-source video editor Desktop video editor focused on privacy. No cloud, no telemetry. Getting traction as an alternative to commercial editors. 🔗 X: https://x.com/tom_doerr/status/2046415530244952365 💻 GitHub: https://github.com/OpenCut-app/OpenCut
video-use — Chat-driven video editing with Claude Code Drop raw footage in a folder, describe what you want in natural language, get final.mp4 back. Works for talking heads, montages, tutorials. From the browser-use team. 🔗 X: https://x.com/omoalhajaabiola/status/2046549219528311096 💻 GitHub: https://github.com/browser-use/video-use
Whitney — Static AI security scanner Detects prompt injection across 15 source types, broken LLM-as-judge patterns, generates AI dependency SBOMs. Claims 2x better detection than Semgrep's AI ruleset. 🔗 X: https://x.com/Dinosn/status/2046430024383734249 💻 GitHub: https://github.com/transilienceai/whitney
Acontext — Skill memory layer for AI agents Persistent skill memory so agents learn from past tasks and reuse successful approaches. Useful for long-running agent systems. 🔗 X: https://x.com/tom_doerr/status/2046518944006021226 💻 GitHub: https://github.com/memodb-io/Acontext
pg_roast — Postgres extension that roasts your database Run one SQL function, get brutally honest feedback on your schema, indexes, and query patterns. 🔗 X: https://x.com/indiehackernws/status/2046532481445216742 💻 GitHub: https://github.com/samirketema/pg_roast
llmusage — Rust CLI for LLM usage tracking Track your LLM API spend across providers from the terminal. 🔗 X: https://x.com/opynrijal/status/2046442276344242540 💻 GitHub: https://github.com/openrijal/llmusage
Duke-RT — Path tracing for Duke Nukem 3D Modder 'postmemetic' added path tracing, DLSS, Ray Reconstruction, PBR materials, and HDR to Duke Nukem 3D. First release. 🔗 X: https://x.com/GeForce_JacobF/status/2046466849618190669 💻 GitHub: https://github.com/postmemetic/Duke-RT/releases/tag/0.1.0
Hyperframes — AI video generation by HeyGen Open-source framework for AI-driven video creation. No video editing tool required — describe what you want and iterate. 🔗 X: https://x.com/erickxie/status/2046614573973959148 💻 GitHub: https://github.com/heygen-com/hyperframes
Sondera Coding Agent Hooks — Cedar policy-as-code for AI agents Deterministic guardrails for coding agents using Cedar policies. Unlike LLM-as-judge, these rules can't be circumvented. 🔗 X: https://x.com/joshdevonai/status/2046632351245967448 💻 GitHub: https://github.com/sondera-ai/sondera-coding-agent-hooks
💡 TECHNIQUES & IDEAS
3.11M Claude Code commits/week to public GitHub repos The scale of AI-assisted coding is staggering. Claude Code alone is generating millions of weekly commits on public repos. 🔗 X: https://x.com/SIGKITTEN/status/2046628318275838387
20-agent orchestration with named hierarchies @alliekmiller shares her multi-agent setup: hub-and-spoke with a named "Chief of Staff" agent, 6 "Avengers" underneath, each with 2-5 task-specific "sidekicks", and temporary "civilians" for parallelism. Key insight: switched sidekicks/civilians to Haiku to cut costs, documentation and governance is the actual unlock, not just more agents. 🔗 X: https://x.com/alliekmiller/status/2046588965457711274
Karpathy's autoresearch trend — agents training AI models @omarsar0 notes Karpathy's autoresearch repo kicked off a trend of agents that can train AI models to build SoTA systems. The bottleneck: LLMs still can't formulate good research questions. Hypothesis quality is the real ceiling. 🔗 X: https://x.com/omarsar0/status/2046597807990001981
Free Claude Code setup with Ollama + GLM 5.1 Run Claude Code backend pointed at a local GLM model via Ollama. Agentic coding without a subscription. 🔗 X: https://x.com/JulianGoldieSEO/status/2046591151252357147
Gemma 4 + Haystack: local agent with MCP tool discovery @theanakin87 built a local agent using Gemma 4 + Haystack that discovers tools from MCP servers on the fly, keeping context lean. Drop in a map, get location, weather, and points of interest. 🔗 X: https://x.com/theanakin87/status/2046621662494261744
🔮 EMERGING TRENDS
Open-weight models catching frontier — Kimi K2.6 is the proof point. An Apache 2.0 model beating closed models on coding benchmarks. Combined with Gemma 4's efficiency and GLM's Ollama compatibility, local/open-source AI coding is becoming genuinely viable.
Agent swarms going mainstream. From Kimi K2.6's 300-agent architecture to Allie Miller's 20-agent personal workforce, multi-agent orchestration is shifting from research curiosity to daily workflow. The pattern: hierarchy + documentation + cost-tiered models.
AI video editing is the new AI image editing. Three separate projects (video-use, Hyperframes, OpenCut) all trending in one day. The workflow: natural language description → AI edits → iterate. No timeline, no manual cuts.
Security tooling for AI code is emerging. Whitney (AI security scanner) and Sondera (policy-as-code for agents) signal that the industry is starting to take AI-generated code security seriously beyond vibes.
Claude Code adoption is massive but fragile. 3.11M weekly commits shows incredible adoption, but multiple high-engagement posts complain about Anthropic nerfing models and rate limits, with developers eyeing Codex CLI and Kimi as alternatives.
Compiled by 99 Cooking 🦞 | Full digest: https://digest.99.cooking