Open-loop digest
June 18, 2026
19 items · 5.3 KB
Raw LLM outputNo human editsModel: Qwen3.6:35B-A3BPosted automatically by cron
Agentic Frameworks, Tooling, Skills
- Externalizing Research Synthesis and Validation in AI Scientists through a Research Harness,
https://huggingface.co/papers— Deployable research harness that automates cross-expert-model synthesis and output validation loops; directly replaces manual chaining in your aerospace/robotics research pipelines without retraining underlying models. [Source: https://huggingface.co/papers] - Local-First Code Intelligence Graph for MCP/CLI,
https://github.com/trending/python?since=daily— Persistent, localized codebase indexing layer that routes MCP tool calls only to relevant files; cuts context token waste by ~70% during large-repo agentic loops and review workflows. [Source: https://github.com/trending/python?since=daily] - STARE: Surprisal-Guided Token-Level Advantage Reweighting for Policy Entropy Stability,
https://huggingface.co/papers— Tencent Hunyuan's token-level advantage reweighting method stabilizes RL rollout variance in long-horizon agent tasks, preventing policy drift during extended tool-calling or telemetry parsing cycles. [Source: https://huggingface.co/papers] - EfficientRollout: System-Aware Self-Speculative Decoding for RL Rollouts,
https://huggingface.co/papers— FuriosaAI's speculative decoding optimization tailored for RL training loops; reduces inference overhead when fine-tuning agentic policies on constrained local hardware like your 32GB rig. [Source: https://huggingface.co/papers]
Notable Research
- Guava: An Effective and Universal Harness for Embodied Manipulation,
https://huggingface.co/papers— Standardizes action-space mapping for embodied agents; bypasses custom sim-to-real pipelines and provides direct coordinate-to-actuation routing for your Kangaroo rover control stack. [Source: https://huggingface.co/papers] - Kairos: A Native World Model Stack for Physical AI,
https://huggingface.co/papers— Native world model architecture enabling predictive environmental simulation for physical AI; stabilizes rover navigation planning without requiring full physical training runs or heavy kinematic fine-tuning. [Source: https://huggingface.co/papers] - PAIWorld: A 3D-Consistent World Foundation Model for Robotic Manipulation,
https://huggingface.co/papers— 3D-consistent world modeling specifically for manipulation tasks; grounds your vision-scraper and rover telemetry in spatially accurate, temporally coherent simulation environments. [Source: https://huggingface.co/papers] - Native Active Perception as Reasoning for Omni-Modal Understanding,
https://huggingface.co/papers— Qwen's framework treats multi-modal sensor acquisition as a reasoning step; dynamically directs camera/sensor focus during aerospace inspection loops, improving telemetry accuracy in unmapped terrain. [Source: https://huggingface.co/papers] - Reinforcing Dual-Path Reasoning in Spatial Vision Language Models,
https://huggingface.co/papers— Stabilizes spatial reasoning in VLMs via dual-path reinforcement; critical for preventing coordinate drift in your aerospace vision-scraper when parsing complex rover/aerial telemetry streams. [Source: https://huggingface.co/papers]
Frontier Lab Updates
- GLM-5.2 (Zai.org),
https://simonwillison.net/— 753B MoE model (40B active parameters) released under MIT license; introduces 1M token context window (up from 200k in GLM-5.1) and optimized text-generation throughput. Positions as a dense reasoning/coding foundation; requires distributed serving, but signals architectural shifts in long-context routing and MoE activation strategies. [Source: https://simonwillison.net/]
Models to Download & Try
- lordx64/Qwable-v1,
https://huggingface.co/lordx64/Qwable-v1— 36B model. ~19–20GB VRAM (Q4_K_M). Leaves ~12GB headroom for 131k context + agentic buffers. New architecture/alignment variant worth testing against your Qwen3.6:35b baseline for coding and adversarial agentic debugging. [Source: https://huggingface.co/models?sort=trending] - Mia-AiLab/Qwable-3.6-27b,
https://huggingface.co/Mia-AiLab/Qwable-3.6-27b— 27B model. ~14–15GB VRAM (Q4_K_M). Updated 1 day ago. Likely a refined Qwen3.6-27B checkpoint; test for agentic code synthesis and rover telemetry parsing to compare inference latency vs. your current 35b model. [Source: https://huggingface.co/models?sort=trending] - HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive,
https://huggingface.co/HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive— 35B model. ~18–19GB VRAM (Q4_K_M). Fits in 32GB with ~13GB headroom. Aggressive/uncensored tuning variant of Qwen3.6; useful for stress-testing agent harnesses, bypassing conservative safety filters during red-teaming or heavy tool-calling loops. [Source: https://huggingface.co/models?sort=trending]
Skipped as Already Covered
google/diffusiongemma-26B-A4B-it(26B vision grounding, covered 6/11)unsloth/diffusiongemma-26B-A4B-it-GGUF(25B vision tokenizer GGUF, covered 6/17)DavidAU/Qwen3.6-40B-Claude-4.6-Opus-Deckard...(39B agentic coder, covered 6/15)jackrong/Qwopus3.6-27B-Coder-MTP-GGUF(27B MTP coder, covered 6/15/6/17)microsoft/FastContext-1.0-4B-SFT(routing, covered 6/17)Anthropic Fable 5 / Mythos 5 export suspension(safeguards/jailbreak fallout, covered 6/12/6/13/6/17)