Tenkai Daily — July 8, 2026
Model Releases
bottlecapai/ThinkingCap-Qwen3.6-27B 🤖 — A multimodal fine-tune of Qwen3.6-27B that aims for token-efficient “thinking” in image-text conversations. If you need a conversational vision model that doesn’t burn through tokens like a bonfire, it’s worth a look — just don’t expect miracles from the marketing flavor.
eric-venti-seeds/Sun-Direction-Lora-Flux2Klein9B — A LoRA for FLUX.2-klein-9B that steers sun direction and lighting in image-to-image generation. Handy if your pipeline keeps producing overcast scenes when you wanted a golden hour, and it’s Apache-2.0 so you can actually use it.
Research Worth Reading
Prompt-to-Paper: Agentic AI System for Bioinformatics 📄 — Tries to fix automated manuscript generation by grounding claims in verifiable literature rather than hallucinated citations. Useful if you’re building agents that write science, less so if you enjoy manually fact-checking LLM footnotes.
From Graphs to Gradients: Physics-Inspired Structural Attribution for Cyber-Physical IoT Systems and Beyond — Proposes a physics-inspired attribution method for explaining behavior in cyber-physical systems. If you need interpretability that goes beyond saliency maps in IoT settings, this might be a refreshing angle.
Foundation Models for Automatic CAD Generation 📄 — Empirical study of using LLMs and VLMs to generate parametric 3D CAD designs from natural language. Could shave time off mechanical design workflows, assuming your specs are cleaner than most real-world ones.
Narrative World Model: Narratology-Grounded Writer Memory for Long-Form Fiction — Adds a narratology-aware memory to track story state (who knows what, when) for long fiction. Another memory architecture, but at least it’s aimed at a concrete multi-hop reasoning problem instead of vague “agent memory.”
FirstResearch: Auditable Question Formation for LLM Scientific Discovery Agents 📄 — Focuses on making the initial research question from discovery agents auditable and traceable. For anyone shipping scientific agents, provenance of the first hypothesis is a real gap this tries to close.
Memory in the Loop: In-Process Retrieval as Extended Working Memory for Language Agents 🔥 — Argues memory should be read/written every step inside the agent loop, not queried once per turn. This challenges the standard RAG-agent pattern and could matter for agents that need tighter state tracking.
AI Dev Tools
MadsLorentzen/ai-job-search 🛠️ — A Claude Code-based framework that evaluates jobs, tailors CVs, writes cover letters, and preps interviews from your profile. Automates the part of job hunting nobody misses, though you’ll still need to show up for the meetings.
iOfficeAI/OfficeCLI — A single-binary, open-source CLI that lets agents read/edit Word, Excel, and PowerPoint without installing Office. If your agent stack needs to manipulate docs without COM nightmares, this is a clean option.
Today’s Synthesis
If you’re building any agent that needs to track state across more than a handful of steps, today’s research drop suggests it’s time to retire the “retrieve once per turn” RAG habit. Memory in the Loop: In-Process Retrieval as Extended Working Memory for Language Agents 🔥 makes the case that memory should be read and written on every step inside the agent loop — and Narrative World Model: Narratology-Grounded Writer Memory for Long-Form Fiction shows a concrete way to do structured state tracking (who-knows-what-when) without boiling the ocean on generic “agent memory.” A practical move: fork your agent loop, add a read/write call to a small structured store at each transition, and steal the explicit state schema idea for your own domain instead of dumping everything into a vector db. If your agent writes science rather than fiction, pair that with FirstResearch: Auditable Question Formation for LLM Scientific Discovery Agents so the opening hypothesis is logged before the loop even spins up. You get traceable provenance and tighter state tracking — two things every post-mortem wishes the system had from day one.