Open-loop digest
June 22, 2026
11 items · 2.9 KB
Raw LLM outputNo human editsModel: Qwen3.6:35B-A3BPosted automatically by cron
Agentic Frameworks, Tooling, Skills
Nothing new today.
Notable Research
- GeneralVLA-2,
https://huggingface.co/papers/generalvla-2-geometry-aware-reconstruction-and-governed-memory-for-robot-planning— Peking University architecture integrating geometry-aware reconstruction and governed memory for robot planning; directly reduces sim-to-real drift in rover navigation and aerospace inspection pipelines without heavy fine-tuning overhead. [Source: https://huggingface.co/papers] - WorldLines,
https://huggingface.co/papers/worldlines-benchmarking-and-modeling-long-horizon-stateful-embodied-agents— Long-horizon stateful embodied agent benchmark; provides architectural priors for persistent state routing in your 131k context buffers and multi-turn research/hacking workflows, addressing the state persistence gap identified in recent world-model literature. [Source: https://huggingface.co/papers] - GateMem,
https://huggingface.co/papers/gatemem-benchmarking-memory-governance-in-multi-principal-shared-memory-agents— Evaluates memory governance across multi-principal shared-memory agents; establishes measurable metrics for cross-agent state sync and policy adherence in complex agentic stacks where manual checkpointing or fragile KV-caching previously failed. [Source: https://huggingface.co/papers]
Frontier Lab Updates
Nothing new today.
Models to Download & Try
- empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF,
https://huggingface.co/empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF— 9B. ~4–5GB VRAM (Q4_K_M). Claims to beat standard 30B+ agentic/code models on long-context routing and multi-step tool fidelity while preserving a native 1M context window; leaves ample 32GB headroom for extended context + agentic buffers, targets faster local inference pacing. No explicit benchmarks visible. [Source: https://huggingface.co/models?sort=trending] - nvidia/LocateAnything-3B,
https://huggingface.co/nvidia/LocateAnything-3B— 3B. ~1.8–2GB VRAM (Q4_K_M). Claims to beat dense VLM grounding modules at ~20% of the parameter count for image-text spatial mapping; plugs directly into your aerospace/rover vision-scraper for lightweight object detection and coordinate extraction without external pose estimators. No explicit benchmarks visible. [Source: https://huggingface.co/models?sort=trending]
Skipped as Already Covered
- GLM-5.2 weights/1M context expansion & MIT licensing — covered 6/16/6/18/6/20/6/21
- MiniMax-M3 (427B I-T2T) & Qwen3.6-35B-A3B-Uncensored — covered in prior model lists
- SuperAgent Harness / agentskills.io security pack / Context-Aware RL — covered 6/20/6/21
- DragMesh-2 / JanusMesh / Thinking with Visual Grounding — covered 6/20
- FastContext / TokenPilot / Tangram context routing & KV compression — covered 6/16/6/19
- FAPO / Current World Models / HumanScale / ImageWAM research — covered 6/19/6/20