Model Releases

NVIDIA Nemotron Labs VoiceChat-11B: Voice Conversation Model — An 11B parameter voice chat model fine-tuned from Nemotron-Nano-9B-v2, backed by five arXiv papers covering the architecture and training. Licensed under OpenMDW-1.1. Built for real-time voice interaction — think voice assistants that don’t sound like they’re reading a script. 🤖

MiniMax-H3 Turbo LoRA: 4-Step Accelerated Inference — Community Turbo LoRA adapter that slashes MiniMax-H3 inference to 4 steps for text-to-video, text-to-audio, and synchronized AV generation. ComfyUI compatible, Apache 2.0. If you’ve been waiting for real-time video-audio gen without the wait, this is it.

MiniMax-H3 Turbo LoRA ComfyUI (Pruned, 4-Step) — Alternative Turbo LoRA with pruned weights, curve-form optimization, and a checkpoint-500 training state. BF16 precision, partial conversion. Same 4-step speed, different tradeoffs. Apache 2.0.

MiniMax-H3 GGUF Quantizations for Local Inference — Community GGUF quants of MiniMax-H3 (via Comfy-Org distro) for running the multimodal video-audio model on consumer hardware. License unspecified, so check before commercial use. 🛠️

Open Source Releases

agentic-hil 0.9.0 - AI-assisted Hardware-in-the-Loop for embedded firmware — HIL tooling that brings AI assistance into embedded firmware dev loops. Simulate, validate, and iteratively test AI models on target hardware. Bridges the gap between “it works in Colab” and “it runs on the MCU.” 🤖

converseek 0.2.7 - Cross-tool session search for AI coding assistants — Search, browse, and export session history across Claude Code, Cursor, Codex, and others. Solves the context fragmentation when you’re juggling multiple AI coding tools. Finally, a unified history. 📄

lorekeep 0.21.1 - Temporal knowledge graph for AI agents via MCP — Time-aware knowledge graph for agent memory using Model Context Protocol. Enables long-term, versioned context across extended interactions. Agents that actually remember what happened last week. 📄

metrik-ai 0.5.0 - AI systems engineering platform for transformer observability — Production observability for transformers: performance profiling, latency analysis, optimization recommendations. If you’re running LLMs in prod and flying blind, this is the dashboard you’ve been building in your head. 🛠️

Research Worth Reading

Woodpecker Distillation: Weak Models Diagnose Reasoning Bugs in Strong Models — Weak probe models identify and patch localized reasoning bugs in strong LLMs. Generates short corrective patches for intermediate steps — targeted repair without full retraining. Turns out many reasoning failures are surprisingly fixable. 📄

QEvict: Recoverable Quantized KV Eviction for Attention-Drift-Robust Long-Context Decoding — Solves the irreversible token eviction problem in KV cache compression. Tokens can be reloaded if later deemed important, making eviction robust to attention drift. Adds quantization for extra memory savings. Long-context decoding just got less lossy. 📄

SearchAuditor: Auditing and Attributing Failures in Long-Horizon Search Agents — Automated framework for diagnosing failures in deep search agents with long web interaction traces. Attributes errors to specific reasoning steps without manual inspection of massive trajectories. Essential for building agents you can actually debug. 📄

Beyond Full-Model Rollback: AuroSFT for Adapter-State Multi-Task Fine-Tuning — Per-task adapter-state rollback instead of full-model rollback for multi-task SFT. Different tasks peak at different training stages; AuroSFT tracks and reverts only relevant adapter params per task. Avoids catastrophic forgetting in heterogeneous task mixes. 📄

Beyond Rotations: AuroOFT for Expressive Quantized Orthogonal Fine-Tuning — Extends QOFT with input-dependent nonlinear corrections via low-rank adapters. QOFT is limited to linear orthogonal transforms; AuroOFT adds expressivity while keeping parameter efficiency for low-bit models. Relevant if you’re deploying adapted quantized models. 📄

Search2Skill: Skill Distillation Beyond Knowledge Boundaries Via Rubric-Based Reinforcement Learning — LLM agents acquire reusable skills beyond parametric knowledge by combining web search with rubric-based RL. Skills distilled from search-augmented trajectories, validated against structured rubrics. Self-evolution in expert domains. 📄

AI Dev Tools

Panniantong/Agent-Reach: Internet access CLI for AI agents (no API fees) — Single binary giving agents read/search access to Twitter, Reddit, YouTube, GitHub, Bilibili, XiaoHongShu without API keys. Works with Claude Code, Cursor, others via MCP. Research-oriented agent workflows just got a free browser. 🛠️

Today’s Synthesis

If you’re shipping agents that run for hours across web searches, tool calls, and multi-step reasoning, you’re flying blind without three things: a memory that survives context windows, a way to pinpoint why it failed at step 47, and production observability that doesn’t require grepping logs. lorekeep gives you a temporal knowledge graph via MCP — agents that remember last week’s decisions, not just the last 8k tokens. Pair it with SearchAuditor to automatically attribute failures to specific reasoning steps in those massive interaction traces, no manual inspection required. Then wire metrik-ai for transformer-level observability: latency profiles, KV cache pressure, token throughput — the dashboard you’d build yourself if you had time. Together, that’s a debuggable, observable agent stack: long-term memory, failure attribution, and runtime telemetry. Start by instrumenting one long-horizon workflow with all three; the first post-mortem will pay for the integration effort. 🛠️📄