Model Releases

  • tencent/Hy-MT2-30B-A3B — 30B MoE translation model covering 30+ languages (zh, en, fr, pt, es, ja, tr, ru, ar, ko, th, it, de, vi, ms, id, tl, hi, pl, cs, nl, km, my, fa, gu, ur, te, mr, he, bn, ta, uk, bo, kk, mn, ug). Sparse A3B architecture keeps inference reasonable. Worth benchmarking if you’re replacing a translation pipeline that currently shuffles between models.

  • circlestone-labs/Anima — Single-file diffusion model, ComfyUI-compatible. “Single-file” is the hook here — version-control a complete workflow without dependency hell. Sparse details on the page, but if you’re into self-contained diffusion setups this is worth a look.

  • Efficient-Large-Model/SANA-WM_bidirectional — Text-to-video and image-to-video with camera control and a world model, bidirectional conditioning. ArXiv:2605.15178. Apache 2.0, so you can actually run it. If you’re experimenting with controllable video generation, put this on the shortlist.

Open Source Releases

  • Rett fra Bonden — Lokal Matfinner — Chrome extension (1,945 installs) indexing 1,433+ Norwegian local-food producers across 368+ cities — farms, shops, REKO rings, markets, dairies, fishers, bakeries. Returns ranked results. Useful if you’re building local-food search in Scandinavia; otherwise a very specific niche.

  • travel-deals-mcp — MCP server letting AI agents search curated resort packages and travel deals in real time. No website crawling, no SEO. 4,060 installs. If you’re building a vacation-booking agent, this is a ready-made data source. If not, it’s a niche MCP.

  • vistoya — MCP server giving agents access to a curated multi-brand fashion catalog. Structured filters, natural language search, similar-item discovery, full product details. 2,020 installs. Targeted at fashion shoppers. Another MCP layer — usefulness depends on your use case.

  • anthropics/claude-code v2.1.149 — Per-category usage breakdown in /usage (skills, subagents, plugins, per-MCP-server costs) and keyboard-scrollable diff view in /diff. Small quality-of-life improvements that make the CLI less painful for long sessions. Not flashy, just less friction.

  • sst/opencode v1.15.9 — Redesigned diff viewer with file tree, returns to previous screen on close, clearer errors for invalid models and missing PTY sessions. Core code review/editing UX polish. If you use Opencode, this is a welcome refresh.

  • awaithumans 0.1.6 — HITL infrastructure for AI agents. Agent calls await_human(), human reviews via Slack/email/dashboard, agent resumes with typed response. Lightweight approach if you need approval gates in your agent workflows.

AI Dev Tools

  • cline/cline CLI v3.0.13 — Loading dialog on session resume so the TUI doesn’t freeze, faster /clear by deferring session creation until next prompt. UX wins for a TUI people actually use daily. Not a feature parade, just less friction.

MCP Servers & Integrations

  • Case Law Search — MCP server for searching 9M+ court opinions and federal dockets. 2,028 installs. If you’re building legal research agents, this is a solid data source. Otherwise, niche.

Tutorials & Guides

  • karpathy/nn-zero-to-hero — Karpathy’s neural networks course repo. From basics to advanced with hands-on exercises. If you want to actually understand what’s under the hood, start here. No shortcuts, no hype.

  • rohitg00/ai-engineering-from-scratch — Builds and ships AI systems from scratch. Practical focus on real-world engineering skills. Good companion to Karpathy’s if you want to go from “I understand the math” to “I can deploy this.”

Today’s Synthesis

If you’re building an AI agent that actually does something useful (not just chat), start with the travel-deals-mcp and vistoya MCP servers — they give you real-time access to resort packages and fashion catalogs with structured filters, so your agent can search across multiple brands without crawling websites. Both MCP servers return structured data, so your agent can filter by price, dates, or style without parsing HTML. Then add awaithumans 0.1.6 to gate any transactional action: the agent searches, presents options, and waits for human approval via Slack or email before proceeding. Pair this with anthropics/claude-code v2.1.149 for a CLI that shows per-MCP-server usage and lets you scroll through diffs without losing context, so you can track costs per data source. The result is an agent that can browse travel and fashion options across multiple brands, stay within budget (thanks to the usage breakdown), and never commit without a human sign-off. The HITL loop means you can test the agent on low-stakes queries before letting it handle real bookings. Practical, not hypothetical. travel-deals-mcp , vistoya , awaithumans , anthropics/claude-code v2.1.149