Preview: Liquid AI's LFM2.5 2.6B, Qwen 3.8 open weights, full-duplex voice, and EU AI Act tools.
Hey — As local hardware constraints force us to think beyond VRAM stacking, non-transformer architectures and deterministic safety gates are quietly redefining how we run agents on-device.
🔥 This Week's Big One
Liquid LFM2.5-2.6B: Non-Transformer Edge Intelligence
Liquid AI released LFM2.5-2.6B, a 2.69B parameter hybrid model that trades standard Transformer self-attention for linear recurrence and state-space layers. Featuring a 128K context window and requiring under 2.5 GB of RAM when quantized, it achieves speeds up to 220 tokens/sec on Apple Silicon and runs comfortably on mobile hardware. It proves that architectural efficiency—rather than raw VRAM capacity—is becoming the primary lever for high-throughput on-device inference. Hugging Face
🧠 Model Releases
- NVIDIA NemotronLabs VoiceChat 11B — A full-duplex conversational voice model combining speech understanding and generation in a unified stack, delivering under 450 ms latency without cascaded TTS pipelines. Hugging Face
- Qwen 3.8-Max & Qwen 3.8-27B — Alibaba announced open-weight releases for its 2.4T MoE flagship alongside a desktop-friendly 27B sibling designed to run within 17 GB VRAM. Qwen AI Blog
- poolside Laguna S 2.1 — A 118B parameter open-weights MoE model featuring a 1-million-token context window tailored for deep agentic coding and workspace-wide refactoring. Hugging Face
🛠 LM Studio / Ollama
- Ollama v0.32.6 — Adds MLX engine performance optimizations for Qwen models on Apple Silicon and aligns streaming output with standard OpenAI API specs. GitHub
- llama.cpp b10276 — Introduces a per-conversation working directory picker supporting cross-platform OS paths alongside locked dependency builds for server security. GitHub
🤖 Agents & Frameworks
- ToolJet ActionRail — An open-source runtime enforcement layer that acts as a deterministic safety gate, verifying proposed agent API actions against systems of record before execution. GitHub
- Nightcrawler — An autonomous red-teaming pentesting agent built by GarageHQ that executes multi-step security checks fully offline on Android smartphones using a local 1.2B model. GitHub
📦 Tools & Repos
- Graphify — A tree-sitter codebase mapping utility that converts repositories into queryable knowledge graphs, paring down agent token context consumption by up to 70x. GitHub
- AIR Blackbox & Agent Audit Trail MCP — Compliance toolsets that record tamper-evident SHA-256 audit trails locally to satisfy EU AI Act record-keeping requirements. GitHub
⚡ Quick Hits
- Mistral Shieldstral-1.0-3B — A 3B policy-adaptive multimodal safety classifier that moderates text and images on-device using natural language guidelines. Hugging Face
Keep your models local and your execution loops grounded.
— Himanshu
