Tenkai Daily — September 12, 2026

Model Releases

• Edge0/Edge0-35B-A3B-preview — MLX-optimized 35B MOE with 4-bit precision, edge prerouter, and SSD-offload for Apple Silicon. Apache 2.0, good for on-device conversational workloads without melting your laptop. 🤖
• openbmb/MiniCPM5-2B-GGUF — GGUF-quantized MiniCPM5-2B for on-device edge AI with long-context, tool-calling, and multimodal abilities. Cites arxiv:2506.07900 and arxiv:2602; reduced footprint without sacrificing core tool-calling.
• nex-agi/Nex-N2.5-Pro — Qwen3.5 MOE with compressed-tensors optimization, endpoints_compatible API. Apache 2.0, text and image-text generation with better token efficiency.

Open Source Releases

• jax-mps 0.10.11.dev834 — JAX backend for Apple Metal Performance Shaders. Accelerates array ops on macOS GPU, bridges JAX and native GPU performance. 🛠️
• agentknit 0.1004.0 — Framework for coding agents against OpenAI-compatible endpoints with spec-driven tools, cache verification, and sandboxing. 0.1004.0 adds context compaction and deterministic execution for lower token usage.
• limatus 0.27.0 — Diagnose-and-steer quality loop for AI-generated content with real-time feedback. 0.27.0 adds enhanced hallucination pattern detection and fine-grained content attribute control.
• savi-sdk 0.15.0 — Observability SDK for LLM operational costs, carbon emissions, and compliance metrics. 0.15.0 adds fine-grained per-request cost attribution and major provider integrations.

Research Worth Reading

• Automating Quadratic Unconstrained Binary Optimization (QUBO) Formulation Generation from Natural Language — Parses natural language into QUBO formulations for combinatorial optimization. Bridges NLP and quantum/hybrid solvers. 📄
• Understanding LoRA Rank Trade-offs in Diffusion Model Fine-Tuning — Controlled study on LoRA rank {2,4,8,16,32} for CIFAR-10 DDPM U-Net fine-tuning. Rank 8 hits the sweet spot on the quality-compute Pareto frontier.
• Towards a Deterministic Math Solver for Clinical Language Models — Clinical LLMs generate Python scripts instead of computing directly; validated interpreter executes. Deterministic arithmetic results for clinical calculations.
• Decoupling Readiness from Release for Tail-Aware Scheduling of Agentic LLM Workflows — Immediate release of ready turns under contention degrades efficiency. Tail-aware scheduling queues ready turns, decoupling readiness from release.
• Probabilistic Focal Search: Accelerating Bounded-Suboptimal Search via Lower-Bound Advancement — Randomized Focal Search variant with probabilistic frontier selection within FOCAL threshold. f_min decreases across expansions for better search efficiency in pathfinding.
• An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics — Post-training and test-time inference design for Nemotron 3 Ultra proof generation. SFT + RL specialist checkpoints, verification/refinement strategies. 🧮

AI Dev Tools

• worktrunk: CLI for Git Worktree Management in Parallel AI Agent Workflows — CLI for Git worktree management in parallel AI agent workflows. Branch isolation and context switching across concurrent agent tasks. 🛠️
• pentagi: Fully Autonomous AI Penetration Testing System — Autonomous AI agent for penetration testing: recon, vuln scanning, exploit generation, automated reporting. Offensive security automation. ⚔️

Today’s Synthesis

When orchestrating parallel AI agent workflows, the instinct to fire ready turns immediately creates contention and wasted compute. Applying the tail-aware scheduling principle from Decoupling Readiness from Release for Tail-Aware Scheduling of Agentic LLM Workflows , you can decouple readiness from release by queuing turns based on actual preparedness rather than mere availability. Pair this with worktrunk ’s Git worktree isolation: each agent instance lives in its own branch, preventing context bleed and enabling independent versioning of prompts, dependencies, and model weights. Then, use agentknit ’s spec-driven tool framework to define exactly which endpoints and functions each worktree can invoke, with cache verification and deterministic execution to keep token usage predictable. The result is a fleet of edge- or cloud-deployed agents—whether running MiniCPM5-2B-GGUF for on-device tool-calling or larger MOE models—where only fully ready agents proceed, ready turns queue under contention, and every spawning event is traced through a controlled spec. You get fewer blind executions, lower overall cost, and a clear upgrade path when models shift.