Model Releases

  • DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic — 🤖 Combines Qwen3.5-9B base with reasoning distillation, SFT/RL/GSPO training pipelines, tool-use and function-calling enhancements, and multiple abliterated/decensored variants, all with endpoint compatibility and multilingual support. The “heretic” tag signals abliterated variants—verify output filters match your use case.

  • Qwen3.5-4B-GLB-3D-Coder — 🛠️ Qwen3.5-4B variant integrating GLB 3D asset format and Godot engine support via ThreeJS, enabling code-aware 3D content generation and visualization with Apache 2.0 license and endpoint compatibility.

Open Source Releases

  • NVKE: NVIDIA Kubernetes Engine for Bare-Metal GPU Clusters — Open-source management plane for deploying and operating GPU-accelerated Kubernetes clusters on bare metal, featuring device plugin integration, hardware health monitoring, and cluster autoscaling optimized for AI/ML workloads. Reduces operational complexity of on-premise GPU infra, though you’ll need bare metal hardware to benefit.

  • Anthropic Claude Code v2.1.319: Auto mode and LLM gateway changes — Changes to auto mode behavior when telemetry/feature flags are off, and LLM gateway call restrictions. New server-side restrictions mean calls can’t check or fall back to the built-in classifier; set CLAUDE_CODE_AUTO_MODE_SERVER=0 to keep using it for Bedrock, Vertex and Foundry (unchanged).

  • shortkit-ml 0.2.1: Calibrated shortcut audits for arbitrary embedding spaces — 📄 Releases a calibrated audit framework to quantify shortcut potential across any encoder and modality, generating localized HTML and PDF reports. Designed for systematic bias and feature leakage detection in ML embedding pipelines.

  • zenagent-py 0.9.1: Minimal Pythonic framework for LLM coding agents — Provides a streaming LLM client and async-generator agent loop for building deterministic, interactive coding agents. Includes a terminal harness for tool use, command execution, and agent introspection in developer workflows.

  • Opencode v2.2.6: Shared MCP connections and GUI themes — Introduced shared MCP connections allowing one pooled connection across projects with identical setup, reducing process count and improving startup speed. GUI plugins can now add color themes and render new slots. Protocol 8 from packages/process-daemon.

  • datasheet-lens 0.5.0: On-demand datasheet retrieval for LLM agents — 📄 Offers an index and runtime lookup system for component datasheets, enabling LLM agents to ground responses in original documentation. Supports factual accuracy and tool-augmented generation in agentic contexts.

Today’s Synthesis

Before you wire a custom coding agent into production, test the new Claude Code auto mode path against your LLM gateway. Anthropic’s Claude Code v2.1.319 changes auto mode behavior when telemetry or feature flags are off, and adds server-side restrictions that stop calls checking or falling back to the built-in classifier, so set CLAUDE_CODE_AUTO_MODE_SERVER=0 if you rely on Bedrock, Vertex, or Foundry. Then use zenagent-py to build a deterministic agent loop around your chosen LLM client, or datasheet-lens to ground responses in actual component datasheets rather than model memory. The concrete play: pin the Claude Code version, disable the server-side auto mode classifier via the env var, and route agent actions through a controlled gateway. Pair that with a retrieval layer like datasheet-lens so tool-augmented generation cites real docs, and use zenagent-py’s streaming client and async-generator loop if you outgrow stock Claude Code. This gives you auditable behavior and factual grounding without waiting on upstream changes.