##Model Releases

nvidia/Qwen3.6-27B-NVFP4 — NVIDIA’s FP4-quantized Qwen3.6-27B via ModelOpt. Drops the model to ~13GB VRAM with minimal quality loss — practical if you’re serving 27B params on a single 24GB GPU. Apache 2.0.

InternScience/Agents-A1 — Qwen3.5 MoE tuned for agent workflows (image-text-to-text, tool use). The arxiv paper claims strong benchmark numbers on agent tasks; your mileage will vary depending on whether your “agent” actually needs vision or just JSON parsing. 🤖

meituan-longcat/LongCat-2.0 — Meituan’s follow-up to their long-context series. Eval results are up but details are thin — check the leaderboard before swapping out your current 128k+ context model.

Open Source Releases

togatoga/karukan — Japanese IME for Linux/macOS with a neural kana-kanji engine. Finally, a local conversion engine that doesn’t require sending keystrokes to Google/Microsoft. If you write Japanese on Linux, this is the one. ⌨️

0xNyk/council-of-high-intelligence — 18 personas (Aristotle, Feynman, Kahneman, Torvalds, etc.) run multi-round deliberation across multiple LLM providers. One command: /council. Gimmicky? Yes. Useful for stress-testing decisions against diverse reasoning styles? Also yes. 🧠

yikart/AiToEarn — “Let’s use AI to Earn!” — the README screams crypto-adjacent hustle culture. No technical docs, just promises of passive income. Hard pass unless you enjoy being the liquidity. 🚩

Research Worth Reading

Constructive Alignment: Governing Preference Dynamics in Human-AI Interaction — Challenges the static-preference assumption in alignment. Argues preferences are constructed through interaction — layered, dynamic, context-dependent. If you’re building preference learning systems, this reframes the problem entirely. 📄

Bounded Morality: Defining the Space of Moral Computation — Formal framework for moral reasoning as bounded computation rather than static theory adherence (deontology/consequentialism/etc.). Treats ethics as a resource-constrained optimization problem. Refreshingly computational. 📄

The MMM Data Model — A Normative Specification for Knowledge Interoperability in a Decentralisable Knowledge Commons — Proposes moving beyond document-centric knowledge representation toward a composable, decentralized data model. Ambitious spec; adoption is the hard part. Worth reading if you work on knowledge graphs or RAG architectures. 📄

Making Failure Safe: A Constrained, Verifiable Agent Framework for Open-Web Data Collection — Addresses the “LLM writes a scraper, it breaks immediately” problem. Constrained generation + verification loop for web agents. Practical if you’re building data pipelines that can’t afford hallucinated selectors. 🛠️

Solution space path planning for supporting en-route air traffic control — Path-planning algorithms vs. operational reality in ATC. The gap between “optimal in simulation” and “usable by controllers” is the story here. Relevant for any domain where algorithms meet human-in-the-loop constraints. 📄

RareDxR1: Autonomous Medical Reasoning for Rare Disease Diagnosis Beyond Human Annotation — Autonomous reasoning for rare disease differential diagnosis. Tackles the “vast search space, sparse annotations” problem with self-supervised reasoning traces. Medical AI that doesn’t just memorize training distributions. 📄

AI Dev Tools

allenai/olmocr — PDF linearization toolkit built for LLM training data prep. Handles layout analysis, reading order, table extraction — the unglamorous plumbing that determines whether your model learns from PDFs or learns garbage. 🛠️

TencentCloud/CubeSandbox — Lightweight, instant-start sandbox for agent code execution. Built for concurrency and security (gVisor/Kata underneath). If you’re running untrusted LLM-generated code at scale, this beats rolling your own Docker timeout wrapper. 🛠️

Today’s Synthesis

Think you can trust an LLM to write a scraper? The council-of-high-intelligence gives you a cheap multi‑persona deliberation loop that can argue over selector strategies. Pair that with CubeSandbox to run the generated code in a lightweight, gVisor‑backed sandbox that kills runaway processes before they hit your CPU. Add the constraints from Making Failure Safe — a verification loop that checks the scraper’s selectors against a known DOM snapshot — and you get a pipeline that still fails fast but never explodes the cluster. The result: a one‑click script that pulls tables out of PDFs (thanks to olmocr for layout prep) while the council debates whether to use XPath or CSS, the sandbox enforces timeouts, and the verifier nips hallucinated selectors in the bud. You can also feed the council a prompt like “extract the table in this PDF” and have it generate a chain‑of‑thought plan that the sandbox executes step‑by‑step, while the verifier logs each selector check for auditability. Deploy it as a service, and you’ll stop watching your logs for “scraping‑related” OOM kills.