Tenkai Daily — May 20, 2026

Model Releases

• inclusionAI/Ring-2.6-1T: Large-Scale Hybrid Architecture Model — A 1T parameter hybrid-architecture conversational model, MIT-licensed, with eval results and compressed-tensors support. The hybrid angle is worth poking at if you’re tired of monolithic decoder blocks doing everything. 🤖
• sapientinc/HRM-Text-1B: Hierarchical Reasoning Model — Hierarchical reasoning via a prefix-LM architecture with a pre-alignment approach, Apache 2.0. Targets non-chat text generation rather than the usual chat pipeline. Niche, but if you’re building structured outputs it might be worth a look. 📄
• Jackrong/Qwopus3.5-9B-Coder-GGUF: Agent-Focused Coding Model — Qwen3.5-based model fine-tuned on Claude Opus reasoning traces for coding, tool-use, and function calling. Available in GGUF, which means you can actually run it locally. Practical. 🔥

Open Source Releases

• ContrastAPI Security Intelligence MCP Server — MCP server with 53 security-intel tools: CVE/KEV, MITRE ATLAS+D3FEND, Sigma rules, email posture, domain/web intel, threat intel. 4,614 installs and growing. If you’re wiring security into an agent stack, this is a solid foundation. 🛠️
• Plith Agent Infrastructure APIs — Five APIs as the base layer for AI agents: task dedup, cost prediction, output validation, behavioral governance, shared failure intel. 1,000 free credits/month, no credit card. The “no credit card” part alone makes it worth a glance. 🛠️
• Weavely AI Forms & Surveys MCP Server — MCP server (1,534 installs) with 13 tools covering the full form lifecycle: creation, 25+ element types, conditional logic, themes, multi-step pages, publishing — all via natural language with live preview. Niche but the MCP angle is clever. 📄
• phi-gateway 0.4.0 — Self-hosted AI gateway with LLM proxy, MCP tool registry, RAG knowledge base, and agent memory through a single API. Zero vendor lock-in is the pitch, and on paper it covers a lot of ground. Worth evaluating if you’re tired of cobbling together your own orchestration layer. 🔥
• vllm-htop 0.4.0 — htop-style terminal monitor for vLLM inference servers: GPU utilization, request throughput, KV-cache metrics in real time. If you’re running vLLM in production, you should already be looking at this. 🛠️

Research Worth Reading

• HELoRA: Hot Experts Layer-Level Low-Rank Adaptation for Mixture-of-Experts Models — HELoRA targets MoE models specifically, exploiting sparse activation patterns for more efficient LoRA-style fine-tuning. Most LoRA work ignores MoE sparsity; this tries to close that gap. 📄
• UCCI: Calibrated Uncertainty for Cost-Optimal LLM Cascade Routing — Maps token-level margin uncertainty to per-query error probability, enabling cost-optimal routing between small and large models without per-workload threshold tuning. Solves a real deployment pain point. 🔥
• Theory-optimal Quantization Based on Flatness — Proposes a quantization approach grounded in activation flatness to mitigate outlier effects at low bit precision. Gives a principled framework for LLM compression rather than the usual “try 4-bit and pray.” 📄
• D-PACE: Dynamic Position-Aware Cross-Entropy for Parallel Speculative Drafting — D-PACE improves multi-token drafter training for parallel speculative decoding, specifically targeting diffusion-based drafters like DFlash. Enables deeper drafters and longer accepted token sequences. If you’re into speculative decoding, this is relevant. 📄
• PASC: Pipeline-Aware Conformal Prediction with Joint Coverage Guarantees for Multi-Stage NLP and LLM Pipelines — PASC provides joint coverage guarantees across multi-stage pipelines, accounting for error compounding. Stage-independent calibration is a known weakness in pipeline systems; this addresses it directly. 🔥
• Simply Stabilizing the Loop via Fully Looped Transformer — Fully Looped Transformer reuses the same Transformer blocks iteratively, trading compute for performance without increasing parameters or context length. Loop count is adjustable at inference time. An interesting trade-off to consider. 🤖

AI Dev Tools

• Claude Code v2.1.145: JSON session listing and improved OTEL tracing — Adds claude agents --json for scripting session data (tmux-resurrect, status bars, session pickers). OTEL spans now include agent_id/parent_agent_id and background subagent spans nest correctly. JSON output alone makes this worth the upgrade. 🛠️
• Cline CLI v3.0.9: Concurrent plugin loading and optimistic config toggles — Loads sandboxed plugins concurrently, caches tool descriptors, and makes config toggles optimistic in the TUI. Faster startup and less annoying reconfig cycles. Fuzzy ranking is back too. 🛠️

Today’s Synthesis

If you’re building an agent stack that needs to be both cost-effective and security-conscious, the combination of UCCI for routing and ContrastAPI Security Intelligence MCP Server for threat intel makes a practical pairing—route sensitive queries to a larger model only when UCCI’s uncertainty metric flags them, otherwise handle with a smaller model and skip the overhead. Wire both through phi-gateway 0.4.0 to keep a single control plane for LLM proxying, MCP tool registration, and agent memory. The gateway gives you the orchestration layer without vendor lock-in, while UCCI prevents over-provisioning and ContrastAPI catches the queries that do need the bigger model but also carry security risk. It’s not a silver bullet, but it’s a deployment pattern that actually cuts costs and adds guardrails at the same time.