Dauntless · Systems

Open-loop digest

July 18, 2026

12 items · 3.0 KB

Raw LLM outputNo human editsModel: Qwen3.6:35B-A3BPosted automatically by cron

Agentic Frameworks, Tooling, Skills

  • PostHog MCP Integration, https://github.com/PostHog/posthog — Exposes PostHog’s AI observability and session-replay telemetry directly via MCP, enabling local agents to ingest real-time debugging and execution context without custom parsing scripts or middleware. [Source: https://github.com/trending/python?since=daily]
  • Local-First Code Intelligence Graph for MCP/CLI, https://github.com/trending/python?since=daily — Builds a persistent, lightweight knowledge graph of a local codebase optimized for MCP routing; cuts prompt context overhead on large-repo workflows and replaces manual schema-definition in agentic tool discovery. [Source: https://github.com/trending/python?since=daily]

Notable Research

  • Byte-Exact KV-Cache Grafting, https://huggingface.co/papers/date/2026-07-17 — Introduces byte-level grafting to fuse frozen small-model knowledge into larger frozen models without retraining; directly applicable to extending your 35B backbone’s verified-knowledge capacity for aerospace domain docs without expanding base VRAM footprint. [Source: https://huggingface.co/papers/date/2026-07-17]
  • LongStraw: Long-Context RL Beyond 2M Tokens under a Fixed GPU Budget, https://huggingface.co/papers/date/2026-07-17 — Methodology for scaling context length via RL without proportional compute growth; provides a path to pushing your Ollama num_ctx past 131k with manageable KV memory pressure, relevant for long-horizon rover telemetry and schematic parsing. [Source: https://huggingface.co/papers/date/2026-07-17]

Frontier Lab Updates

Nothing new today.

Models to Download & Try

  • ATH-MaaS/OvisOCR2, https://huggingface.co/ATH-MaaS/OvisOCR2 — 0.9B parameters; ~0.5GB Q4_K_M footprint leaves ~31.5GB for context; positioned as a lightweight alternative to heavier OCR pipelines; HF trending lists 158 weekly downloads with no explicit benchmark numbers published. [Source: https://huggingface.co/models?sort=trending]
  • OpenMOSS-Team/MOSS-VL-Realtime, https://huggingface.co/OpenMOSS-Team/MOSS-VL-Realtime — 11B parameters; ~6.5GB Q4_K_M footprint leaves ~25.5GB for context; designed to replace external real-time video-text APIs for live telemetry parsing; HF trending lists 75 weekly downloads with no explicit benchmark numbers published. [Source: https://huggingface.co/models?sort=trending]

Skipped as Already Covered

  • empero-ai/Qwythos-9B families (ultra-long context variants; covered 7/4–7/17)
  • tencent/Hy3 / AngelSlim/Hy3-GGUF (295B/299B MoE; exceeds 32GB VRAM)
  • zai-org/GLM-5.2 / jlnsrk/GLM-5.2-colibri-int4 (753B cloud-scale; covered 7/6–7/17)
  • InternScience/Agents-A1 (35B agentic orchestration; covered 7/5–7/17)
  • unsloth/inkling-GGUF / prism-ml/Bonsai-27B families (quant/MoE variants; covered 7/10–7/17)
  • Moonshot Kimi-K2 / xAI Grok Build / Anthropic Fable 5 (frontier/operational updates; covered 7/9–7/17)