Open Dreamer ships the full recipe to train a Dreamer 4 world model
The JAX/Flax pipeline trains a playable world model on one GPU — alongside Sakana's security-tuned Fugu endpoint and HKUDS's self-evolving agent framework.
Copy markdown
Train a playable world model on a single GPU
Open Dreamer (from the Reactor team) open-sources a JAX/Flax reproduction of the Dreamer 4 pipeline — causal video tokenizer, a 1.6B action-conditioned dynamics model, rollout and FVD scoring — with the full training recipe: Muon optimizer, 200K steps, ~57% MFU on a single B200. No weights ship, but a browser demo already toggles a real-time Minecraft between “Game” and “Dream,” and you can retrain the pipeline on your own video.
Sakana's Fugu-Cyber: one API that orchestrates security specialists
Sakana added a security-tuned third endpoint to its Fugu orchestrator — a multi-agent system you call like a single model, with a verifier that validates vulnerabilities before it proposes patches. It self-reports 86.9% on CyberGym (vs GPT-5.5-Cyber's 85.6%) and 72.1% on CTI-REALM, but access is gated: Token Plans at $20/$100/$200, $6/$36 per million in/out tokens, no weights, and nothing in the EU/EEA.
OpenSpace: agents that keep the skills they learn
HKUDS open-sourced OpenSpace, a Python framework for agents that write, store, and reuse their own skills — persisted in SQLite as versioned SKILL.md files with lineage tracking, discovered automatically on later runs, and callable through MCP servers. It runs on Python 3.12+ with your own Anthropic or OpenAI key (Claude Sonnet 4.5 by default), so an agent's work on one task compounds into the next instead of starting cold.