AI Daily — 2026-10-13

English 中文

Microsoft BitNet released 1-bit embedding models and VibeASR, while GroundingPI and AgentSWE target physical intelligence and reliable coding agents.

Covering 23 AI news items

🔥 Top Stories

1. Microsoft BitNet Releases 1-Bit Embedding Models and VibeASR Engine

Microsoft’s BitNet project has released VibeASR.cpp, a real-time multilingual ASR inference engine that uses BitNet I2_S quantization on CPUs, along with the first 1-bit embedding models at 0.6B and 270M scales. The updates promise lossless 2-bit-per-weight inference and significant embedding-prefill speedups over F16. This pushes ultra-low-bit inference closer to practical on-device speech and retrieval workloads. Source-github

2. Liquid AI Releases LFM2.5-24B-A2B On-Device Reasoning Model

Liquid AI has launched LFM2.5-24B-A2B, the largest and final model in its LFM2.5 family, with 24B total parameters but only 2.45B active parameters. It targets fast, private on-device agents and reasoning, claiming competitive agentic performance with Qwen3.6-35B-A3B and up to 1.7x faster on-device inference. GGUF weights and a free demo are available, making it a notable entry for local reasoning deployments. Source-reddit

3. AgentSWE Benchmark Tests Whether Coding Agents Build Reliable Agents

Researchers introduced AgentSWE, a benchmark for agent software engineering that spans three stages of the agent development lifecycle. It tests whether coding agents can build agents that faithfully serve user requirements rather than merely appear to. No existing benchmark previously evaluated this assumption, making AgentSWE an important step for coding-agent reliability and safety. Source-huggingface

📰 Featured

Multimodal & Interpretability

GroundingPI: 4B Grounding Foundation Model for Physical Intelligence — A 4B grounding model for precise object grounding, including clutter and tiny objects, that emits quantized points and boxes suitable for closed-loop control and VLA/world-action models. Source-huggingface
S-Space Reveals Spatial Coordinate Representations in Multimodal Language Models — Researchers identify a linear subspace where object-token representations encode horizontal, vertical, and distance coordinates, suggesting intrinsic spatial representations in multimodal LLMs. Source-huggingface

Agentic Reasoning & Evaluation

LLMs Keep Thinking Even When Told Not To, Study Finds — A new paper shows that disabled thinking modes and absent long traces are unreliable proxies for “no-thinking” behavior, since models may still reason or produce filler. Source-huggingface
Scaling Agentic Data for Long-Horizon Terminal Intelligence — The work argues short trajectories miss evidence accumulation, exploration, feedback-driven strategy adjustment, and deliverable generation, motivating high-quality long-horizon agentic datasets. Source-huggingface

Open Source Tools & Self-Hosting

Open WebUI: Self-Hosted AI Interface for Ollama and OpenAI APIs — Open WebUI provides an open-source, self-hosted interface for local and cloud models via Ollama and any OpenAI-compatible API, deployable through pip, Docker, or Kubernetes. Source-github
Open-Source Project Offers Free Access to Multiple AI Coding Agents — free-claude-code aggregates 61 ToS-friendly providers to offer 1.3B+ free tokens monthly across Claude Code, Codex, Pi, OpenCode, VS Code Chat, and other coding agents. Source-github
Company Serves 40 People with 4x RTX 6000 Blackwell Local AI Stack — A 4U server with four RTX 6000 Blackwell MaxQ GPUs runs Proxmox, LXC containers, vLLM-served DeepSeek V4 Flash, Qwen 3.8 27B Q8, and K2-Horizon 7B Q6 alongside n8n, OpenWebUI, Whisper, ComfyUI, and Tavily. Source-reddit

⚡ Quick Bites

Developer Converts Qwen2.5 Dense Models to MoE, Releases GGUF — A developer converted Qwen2.5 dense models into MoE variants and released GGUF weights for local inference. Source-reddit
Jailbroken PS5 GPU Runs Qwen 27B at ~20 Tokens/s — Hacked PS5 hardware reportedly runs Qwen 27B at around 20 tokens/s, highlighting unconventional local inference paths. Source-reddit
Tencent Open-Sources Agent Memory Running Locally on SQLite — Tencent released an agent memory system that runs fully locally on SQLite. Source-reddit
Solo Developer Launches SteerClaw Mobile for Local LLM Agents on Phone — SteerClaw Mobile brings local LLM agent workflows to phones. Source-reddit
RTX 3090 FE VRAM Overheats During LLM Inference, Power Cap Recommended — Users report RTX 3090 FE VRAM overheating under LLM loads, with power capping advised. Source-reddit
AnythingLLM Moves from Ollama to Lemonade and llama.cpp — AnythingLLM is switching its local backend from Ollama to Lemonade and llama.cpp. Source-reddit
Sipeed Announces MaixStation, Comparable to Strix Halo and DGX Spark — Sipeed’s MaixStation targets local AI performance comparable to Strix Halo and DGX Spark. Source-reddit
GPT-2 too dangerous to release: 2019 Guardian article resurfaces — A 2019 Guardian article about GPT-2 being “too dangerous to release” resurfaced amid current AI debates. Source-reddit
Anthropic’s Distillation Defenses Research Called Free Publicity for GLM — Community discussion argues Anthropic’s distillation-defense research inadvertently promotes GLM models. Source-reddit
Reddit User Proposes AI-Assisted Open Source CUDA Reverse-Engineering — A Reddit user floated using AI to assist open-source CUDA reverse-engineering efforts. Source-reddit
User Buys 2x Nvidia DGX Sparks, Seeks GLM Recipes — A user acquired two Nvidia DGX Sparks and is asking the community for GLM deployment recipes. Source-reddit
Anthropic Said Gemma4 26B Runs on 14 Smart TVs — A Reddit thread cites an Anthropic claim that Gemma4 26B runs across 14 smart TVs, spurring local-AI jokes. Source-reddit
Reddit Post Jokes LocalAI Gets Best Advertisement — A Reddit post jokes that LocalAI received its best advertisement from recent AI industry drama. Source-reddit

Generated by AI News Agent | 2026-10-13