Tenkai Daily — March 16, 2026

Open Source Releases

• Qwen3.5-9B Reasoning Distilled Model — A distilled version of Qwen3.5-9B fine-tuned for reasoning, using unsloth. A practical example of compressing and adapting a model for specific conversational tasks without starting from scratch.
• vmlx 1.0.0 — A library for running AI inference locally on Apple Silicon, handling text, image, video, and audio. Useful if you want to prototype or deploy on-device AI without cloud dependencies.
• Qwen 2.5 VL 7B with FP8 Scaling — A quantized vision-language model using FP8, aiming to reduce memory footprint and speed up inference. A solid option for deploying multimodal models where resources are tight.

Research Worth Reading

• ToolTree: Efficient LLM Agent Tool Planning via Dual-Feedback Monte Carlo Tree Search and Bidirectional Pruning — Uses MCTS with dual-feedback and pruning to improve tool planning for LLM agents, specifically tackling inter-tool dependencies. Relevant if you’re building complex agents that need to sequence tool calls intelligently.
• Structured Distillation for Personalized Agent Memory: 11x Token Reduction with Retrieval Preservation — Distills conversation history into a compact retrieval layer, cutting token usage by 11x while keeping retrieval performance. Directly addresses the cost and latency problem of long-context agent memory.
• Efficient Reasoning with Balanced Thinking — Tries to find the sweet spot between overthinking and underthinking in reasoning models, aiming for better accuracy with less compute. A pragmatic look at making reasoning models more efficient.
• AI Planning Framework for LLM-Based Web Agents — Frames web tasks as sequential decision-making for LLM agents, adding structure for better diagnosis and performance. Could help make your web-scraping or automation agents more reliable and debuggable.

AI Dev Tools

• AI Planning Framework for LLM-Based Web Agents — This paper’s framework could be implemented as a dev tool or library for structuring agent tasks, moving beyond prompt engineering to more formal planning.

Today’s Synthesis

The common thread today is making AI agents more efficient and capable without just throwing more compute at them. ToolTree tackles the planning problem, giving agents a smarter way to sequence tool use. But a smart plan is useless if the agent’s memory is bloated and slow; that’s where Structured Distillation comes in, slashing the token cost of remembering past interactions. Meanwhile, Efficient Reasoning with Balanced Thinking addresses the core reasoning loop itself, preventing the model from wasting cycles on unnecessary deliberation. For an engineer, this suggests a practical stack: use structured distillation to manage memory, apply balanced thinking to the core reasoning step, and then layer on something like ToolTree for complex, multi-step planning. It’s a blueprint for building agents that are cheaper to run and more reliable in their actions.