Model Releases

Microsoft releases Fara1.5-27B computer use agent — A multimodal web agent fine-tuned from Qwen3.5-27B for browser automation and computer use tasks. MIT licensed with endpoint compatibility, if you’re building something that needs to click around a browser without human supervision.

Baseten releases GLM-5.2-Vision-NVFP4 quantized vision model — NVFP4 quantization via ModelOpt for SGLang deployment, keeping multimodal vision-language capabilities at 8-bit precision. MIT licensed. Useful if you’re serving vision models and care about throughput over maximal accuracy.

Open Source Releases

agent-inbox 0.24.0 - SQLite mailbox for local LLM agent communication — Local-first messaging infrastructure using SQLite as a mailbox. HTTP API, ActivityStreams model, and stdio MCP server for agent-to-agent comms without external dependencies. Finally, a way to make your agents talk to each other without spinning up Kafka.

helen-lang 1.29.14 - Prompt-first agent programming language — Another agent orchestration framework positioning itself against LangChain, CrewAI, and AutoGen. Prompt-first design, built-in LLM primitives, bilingual support, automatic context management. The “prompt-first” framing is either refreshing or just marketing — your call.

privaite 0.4.1 - Self-hosted LLM proxy with reversible PII redaction — Drop-in proxy that intercepts LLM calls and reversibly redacts PII from prompts, tool arguments, and multimodal content. Privacy compliance without touching application code. Clever approach — redact before send, restore after receive.

Research Worth Reading

DeepLens Diagnosis Agent: Agentic Workflow Design Lets a Small Reasoning Model Compete with Frontier LLMs — Five-stage agentic pipeline (fact extraction, knowledge consultation, differential analysis, diagnosis selection, explanation) enables a small reasoning model to match frontier LLM performance on medical diagnosis. Workflow design > model size, apparently.

Semalith v1.4: A Calibrated 184M Safety Classifier Achieving State-of-the-Art Prompt-Injection Detection at 44x Fewer Parameters than Llama-Guard-3-8B — 184M DeBERTa-v3-base classifier handling prompt injection, regulatory compliance, and general harm detection in one pass. SOTA prompt-injection detection with 44x fewer parameters. Small models doing heavy lifting.

DomainPilot: Domain-Level Loss-Guided Two-Stage Data Mixture Optimization for Efficient Language Model Fine-Tuning — Two-stage data mixture optimization using domain-level loss guidance for terabyte-scale fine-tuning. Avoids O(N) per-sample selection costs of dynamic scheduling. If you’re mixing datasets at scale, this might save you compute.

Same Question, Different Answers: Evaluating LLM Reliability Beyond Accuracy — Studies LLM reliability under meaning-preserving paraphrases across factual QA and math reasoning. High benchmark accuracy ≠ consistent answers to semantically equivalent questions. Your eval suite probably misses this.

MIITA: Memory-Induced Inference-Time Adaptation for Continual Learning with Small Language Models — Memory-based inference-time adaptation for continual learning in small LMs, decoupling knowledge retention from parameter updates to prevent catastrophic forgetting. For when you can’t afford full fine-tuning runs.

SF-AMS: Strategic Forgetting for Structured Memory in LLM Agent — Memory management framework modeling long-term importance of memory units, strategically forgetting low-utility info to maintain compact, high-utility context for multi-step reasoning. Because infinite context isn’t a strategy.

AI Dev Tools

moeru-ai/airi: Self-hosted multimodal AI companion with game-playing — Self-hosted AI companion with realtime voice chat that plays Minecraft and Factorio. Web, macOS, Windows support. A practical demo of LLMs integrated with game environments and voice interfaces for autonomous agents. Also, it plays Factorio — that’s either impressive or concerning depending on your factory layout. 🤖

Today’s Synthesis

The agent stack is finally maturing past “wrap an LLM in a loop and pray.” helen-lang gives you a prompt-first orchestration layer with built-in context management, agent-inbox handles inter-agent messaging over SQLite (no Kafka, no cloud), and SF-AMS solves the context explosion problem with strategic forgetting — keeping only high-utility memories for multi-step reasoning. Wire them together: helen-lang defines the agent graph, agent-inbox passes structured messages between nodes, SF-AMS prunes each agent’s working memory at step boundaries. You get a local-first, observable multi-agent system that doesn’t balloon context windows or require a platform team. Add privaite as a proxy layer if any agent touches PII. This isn’t a demo — it’s a deployable architecture for agents that actually finish tasks. 🛠️