Dauntless · Systems

Open-loop digest

July 20, 2026

10 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

  • Xiaomi-Robotics-1: Scaling Vision-Language-Action Models with over 100K Hours of Real-World Trajectories, https://huggingface.co/papers — Demonstrates policy performance scaling strictly through accumulation of real-world teleoperated trajectories rather than sim-to-real distillation; provides a directly deployable methodology for expanding long-horizon state generalization and tactile feedback loops in your Kangaroo rover without synthetic domain gap calibration. [Source: https://huggingface.co/papers]
  • Recursive Harness Self-Improvement, https://huggingface.co/papers — Introduces a closed-loop execution harness that automatically generates, critiques, and retrains agent behavior from its own failure/success traces; replaces static prompt engineering or external reward-model scaffolding in local agentic stacks, enabling continuous task-completion improvement on complex multi-step pipelines without manual iteration. [Source: https://huggingface.co/papers]

Frontier Lab Updates

Nothing new today.

Models to Download & Try

  • DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF, https://huggingface.co/DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF — 27B parameters; ~16.5GB Q4_K_M footprint leaves ~15.5GB for context (~78k tokens at 16-bit KV). Integrates Fable routing, MTP, and Hermes alignment into a single GGUF; claims superior agentic tool-following and instruction fidelity over base Qwen 3.6:27B configurations for complex rover planning loops. No explicit benchmark numbers visible on trending page. [Source: https://huggingface.co/DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF]
  • nvidia/Nemotron-3-Embed-1B-BF16, https://huggingface.co/nvidia/Nemotron-3-Embed-1B-BF16 — 1B parameters; ~2GB BF16 footprint leaves ~30GB for context. Purpose-built dense embedding backbone to replace heavy cross-encoders in local RAG pipelines; enables high-fidelity semantic routing and vector indexing for your vision-scraper/aerospace documentation without expanding primary model VRAM overhead. No explicit benchmark numbers visible on trending page. [Source: https://huggingface.co/nvidia/Nemotron-3-Embed-1B-BF16]

Skipped as Already Covered

  • thinkingmachines/Inkling / moonshotai/Kimi-K2 family (frontier/weights; covered 7/17–7/19)
  • prism-ml/Bonsai-27B-gguf / Ternary-Bonsai-27B-mlx-2bit (quant/MoE variants; covered 7/18–7/19)
  • empero-ai/Qwythos-9B families (ultra-long context; covered 7/17–7/19)
  • bottlecapai/ThinkingCap-Qwen3.6-27B / prior Qwen3.6 agentic routing variants (NVFP4/MTP; covered 7/18–7/19)
  • zai-org/GLM-5.2 / AngelSlim/Hy3-GGUF (cloud/MoE scale; covered 7/17–7/19)
  • xAI Grok Build / Anthropic Fable 5 subscription & safety policy updates (covered 7/16–7/19)