Model Releases 🤖

Unsloth GGUF Quantizations for DeepSeek-V4-Flash and Kimi-K3 — Unsloth dropped imatrix-optimized GGUFs for both models, so you can run them locally without melting your GPU. Base model refs and inference server endpoints included — plug into llama.cpp, vLLM, or whatever you’re using and go.

Open Source Releases 🛠️

autogluon 1.5.1 beta — Amazon’s AutoML kitchen sink gets a beta refresh: tabular, multimodal, time-series, and feature engineering under one API. Better ensembling, distributed training, updated model zoos. If you’re still hand-tuning XGBoost hyperparameters in 2024, this is either a lifeline or an existential crisis.

sst/opencode v1.18.11 — Fixed MCP SSE reconnect loops that’d hang after server errors, plus support for interleaved reasoning fields (reasoning_text, custom names). Desktop app now opens links in your actual browser instead of trapping them in-app. Small fixes, but the kind that save you 45 minutes of “why is it stuck” debugging.

agent-airlock 0.8.62 — Deny-by-default contract layer for agent tool calls. Pydantic validation, ghost-argument stripping, self-healing retries. Sits in-process for MCP servers and agent frameworks — keeps your LLMs from hallucinating malformed tool invocations or worse, malicious ones.

bentoworks 0.9.4 — Kernel-enforced sandbox for AI agents, no containers or VMs required. Deny-by-default, lightweight isolation for untrusted agent code. If you’re running agents that execute arbitrary code and sleeping fine, you’re either brave or haven’t been pwned yet.

bergson 0.17.1 — Neural network memory tracing via data attribution. Figure out which training samples actually influence a given prediction. Useful for debugging, interpretability, and catching data quality issues before they hit prod. Also good for explaining to your manager why the model did that.

agent-audit-kit 0.3.66 — Security scanner for MCP-connected agent pipelines. Catches vulns, misconfigs, and policy violations in tool definitions, permission scopes, data flow. Drop it in CI and stop deploying agents that can read your .env file.

AI Dev Tools ⚙️

cline/cline SDK v0.0.69 — Ollama timeout bumped from 30s to 5 minutes for cold model loads. Unreachable servers still fail fast, requests stay cancelable, explicit timeouts still respected. Empty responses (no text, reasoning, or tool calls) now handled gracefully instead of silently breaking your flow.

Today’s Synthesis

If you’re shipping agents that touch tools, data, or — god forbid — a shell, you’ve got a three-layer security problem: what the model tries to call, where the code actually runs, and what slips through review. agent-airlock sits at the MCP boundary and enforces Pydantic contracts on every tool invocation — stripping ghost args, retrying malformed calls, and keeping hallucinated payloads from reaching your functions. bentoworks handles the runtime side: kernel-enforced sandboxing without containers, so untrusted agent code can’t escape to your host filesystem or network. Drop agent-audit-kit into CI and it scans your entire MCP pipeline — tool definitions, permission scopes, data flows — for vulns and policy violations before they hit prod. Wire them together: airlock validates at the gateway, bentoworks isolates at the kernel, audit-kit catches drift in CI. You get defense in depth without the container orchestration tax. Start by adding airlock to your MCP server config and audit-kit to your GitHub Actions; bentoworks slots in when you need to run arbitrary agent-generated code.