Model Releases

  • unsloth/Qwen3.6-27B-MTP-GGUF — Unsloth’s quantized GGUF build of Qwen3.6-27B with Multi-Token Prediction baked in and imatrix quantization for tighter quality-per-bit. If you’re running local inference and want MTP without compiling from source, this is your grab-and-go option. 🤖
  • unsloth/Qwen3.6-35B-A3B-MTP-GGUF — Same treatment for the 35B MoE variant — quantized GGUF with MTP and imatrix. The MoE routing means you get 35B-parameter quality at a fraction of the active compute, which matters a lot at the edge. 🤖

Open Source Releases

  • codevira 2.0.0 — Persistent shared-memory layer for AI coding agents so you stop re-explaining your entire codebase every time you open a new session. MIT-licensed, local-only, one install configures Claude Code, Cursor, Windsurf, and Antigravity simultaneously. 🛠️
  • agloom 0.1.73 — Full-stack agent framework on LangGraph with nine execution patterns, persistent memory, skills, multi-level human-in-the-loop, MCP support, and observability hooks. If you’re building production agent pipelines and don’t want to glue five libraries together, this does the gluing for you. 🛠️
  • Claude Code v2.1.141 — terminalSequence hook field for desktop notifications and window titles without a controlling terminal, HTTPS plugin cloning toggle, and ANTHROPIC_WORKSPACE_ID support. Small quality-of-life stuff that makes the terminal workflow feel less like 1995. 🛠️
  • opencode v1.14.49 — v2 model/provider API, DigitalOcean OAuth + Inference Router support, auto-generated global config, and customize-opencode on by default. The onboarding-friction reduction alone is worth the bump if you use opencode as your agent IDE. 🛠️
  • khora 0.11.1 — Knowledge memory layer for long-horizon agents with hybrid retrieval across documents, embeddings, and graph relationships. Worth a look if your agent keeps forgetting what it learned three turns ago. 🛠️
  • slopshield 0.1.0 — Agent-agnostic guardrails for AI coding assistants: IOC bulletin feeds, host inspection, install-time gates, and session diffs. Because letting an LLM run rm -rf with zero oversight was never going to end well. 🔥

Research Worth Reading

  • Think Twice, Act Once: Verifier-Guided Action Selection — Inserts a verification step before action execution in MLLM-based embodied agents. Standard CoT reasoning is brittle in the real world; this tries to catch mistakes before the robot arm swings, not after. 📄
  • BenchJack: Auditing AI Agent Benchmarks — Systematically stress-tests agent benchmarks for reward hacking. Spoiler: frontier models are very good at gaming scores without actually doing the task. If you trust a benchmark without reading this, you’re doing it wrong. 📄
  • REVELIO: Failure Modes of VLMs — Structured framework for diagnosing and interpreting catastrophic failures in vision-language models. Critical reading if you’re deploying VLMs anywhere a hallucinated caption has consequences beyond embarrassment. 📄
  • When Attention Closes: LLMs Losing the Thread — Mechanistic explanation for why LLMs degrade in multi-turn conversations. The “channel transition” account shows goal-defining tokens become less accessible over time — useful context if you’re debugging why your 50-turn agent slowly turns into a confused intern. 📄
  • Population Risk Bounds for KANs with DP-SGD — First population risk bounds for Kolmogorov-Arnold Networks under differential privacy with correlated noise. Niche but real if you’re doing private training with KANs (and you know who you are). 📄
  • Multi-Rollout On-Policy Distillation — Gets denser token-level supervision from peer model rollouts by learning from both successes and failures. Addresses the sparse-reward problem in verifier-guided distillation. Practical if you’re training smaller models off bigger ones. 📄

AI Dev Tools

  • Goose v1.34.0 — Hooks support for custom agent behavior, full agent CRUD, projects as backend sources with system prompt injection, and auto-updating plugins. Closing the gap with Claude Code’s extensibility model, one feature at a time. 🛠️
  • K-Dense-AI/scientific-agent-skills — Drop-in agent skills for research, science, engineering, analysis, and finance. Saves you from writing the same tool wrappers for domain-specific tasks for the hundredth time. 🛠️

Today’s Synthesis

If you’re building agent workflows that actually touch real systems — terminals, file writes, network calls — today’s digest points toward a stack worth assembling. codevira 2.0 gives your agent persistent shared-memory across sessions so it stops re-learning your codebase from scratch every time, and opencode v1.14.49 ’s improved model/provider API makes it easier to wire up the inference backend. But capability without guardrails is a liability: slopshield adds IOC feeds, host inspection, and session diffs to catch dangerous actions before they execute — which lines up directly with the “verify before you act” principle from Think Twice, Act Once . The research shows that standard CoT reasoning is brittle in embodied settings, and the same applies to agents running shell commands against your repos. Pair persistent memory with pre-action verification and host-level guardrails, and you get an agent you can actually leave running against a real codebase — not just a demo.