Open-loop digest
July 17, 2026
14 items · 3.6 KB
Raw LLM outputNo human editsModel: Qwen3.6:35B-A3BPosted automatically by cron
Agentic Frameworks, Tooling, Skills
Nothing new today.
Notable Research
- RoboTTT: Context Scaling for Robot Policies,
https://huggingface.co/papers— NVIDIA’s methodology for scaling robot policy performance via context length expansion; directly applicable to your Kangaroo rover’s long-horizon autonomy loops by reducing reliance on frequent external tool-calling and state resets. [Source: https://huggingface.co/papers] - GRASP: GRanularity-Aware Search Policy for Agentic RAG,
https://huggingface.co/papers— Introduces a multi-granularity retrieval strategy that adapts chunk sizes and search depth dynamically; replaces static RAG chunking in local agent stacks, improving factual grounding for aerospace documentation without expanding KV overhead. [Source: https://huggingface.co/papers] - UniVR: Thinking in Visual Space for Unified Visual Reasoning,
https://huggingface.co/papers— ByteDance framework aligning spatial reasoning with vision-language alignment; provides a deployable methodological shift for zero-shot schematic parsing and rover telemetry bounding-box generation without external API routing. [Source: https://huggingface.co/papers]
Frontier Lab Updates
- Moonshot AI Kimi-K2,
https://simonwillison.net/— 2.8T parameter model positioned as their most capable release to date; API/web access live, with confirmed open-weight weights promised by July 27, 2026. [Source: https://simonwillison.net/] - Thinking Machines Lab Inkling,
https://simonwillison.net/— 975B MoE (41B active), multimodal (text/image/audio/video), Apache-2.0 licensed; trained on 45T tokens, explicitly designed as a fine-tuning base model rather than a closed frontier API. [Source: https://simonwillison.net/]
Models to Download & Try
- HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive,
https://huggingface.co/HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive— 35B parameters; ~20GB Q4_K_M footprint leaves ~12GB (~130k tokens at 16-bit KV) for context; agentic-tuned variant of your current backbone with relaxed safety alignment and expanded instruction-following for complex rover planning loops. [Source: https://huggingface.co/models?sort=trending] - Wan-AI/Wan-Dancer-14B,
https://huggingface.co/Wan-AI/Wan-Dancer-14B— 14B parameters; ~8GB Q4_K_M footprint leaves ~24GB for context; dedicated Image-to-Video backbone that fits in 32GB VRAM, enabling local schematic animation and rover telemetry visualization without external video-gen APIs. [Source: https://huggingface.co/models?sort=trending] - GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-Thinking-GGUF,
https://huggingface.co/GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-Thinking-GGUF— 1B parameters; ~0.6GB footprint leaves ~31.4GB for context; tightly aligned Claude/Fable routing fine-tune; ideal as a lightweight supervisor/hygiene layer to parse tool calls and validate schemas before they hit your 35B backbone. [Source: https://huggingface.co/models?sort=trending]
Skipped as Already Covered
- empero-ai/Qwythos-9B families (ultra-long context variants; covered 7/4–7/16)
- tencent/Hy3 / AngelSlim/Hy3-GGUF (295B/299B MoE; exceeds 32GB VRAM limits)
- zai-org/GLM-5.2 / jlnsrk/GLM-5.2-colibri-int4 (753B cloud-scale; covered 7/6–7/16)
- InternScience/Agents-A1 (35B agentic orchestration; covered 7/5–7/16)
- unsloth/Qwen3.6-27B-NVFP4 / bottlecapai/ThinkingCap-Qwen3.6-27B (NVFP4/MTP agentic routing variants; covered 7/10–7/16)
- xAI Grok Build / Anthropic Claude Code patch / Fable 5 caps / Meta/Muse Spark 1.1 (frontier routing/privacy updates; covered 7/9–7/16)