local ai

Are -mlx variants equivalent to -nvfp4 for gemma4 and qwen3.6?

Ollama has added MLX support for Qwen3.5 and Gemma 4 , with both MLX and NVFP4 quantization variants available for these models. There have been issues with broken Qwen3.6 NVFP4 generation on Ollama ,

local-air-ollama

Content type

All Blog X Post Paper YouTube Reddit GitHub

Clear filters

AtomicMotion: Learning Human Motion From Different Human Parts

DGX agent

arXiv:2605.22631v1 Announce Type: new Abstract: Accurately reconstructing full-body poses from sparse head and hand trajectories is a foundational challenge for immersive AR/VR telepresence. Current m

local-aiarxiv-cs-cv

22 May 2026

Auction-Consensus Algorithm with Learned Bidding Scheme for Multi-Robot Systems

arXiv:2605.21932v1 Announce Type: new Abstract: Multi-Robot Task Allocation (MRTA) is a central challenge in decentralized multi-agent systems, where teams of robots must cooperatively assign and exec

local-aiarxiv-cs-ro

b9276

llama.cpp build b9276 introduces support for hybrid DNA tokenization with new pre-type dispatching and tokenizer implementations, alongside fixes for VRAM leaks in Multi-Token Prediction (MTP) models

b9283

Release b9283 of llama.cpp includes updates to the BPE tokenizer with new pre-type dispatch and Carbon-3B DNA tokenizer support, along with refactoring of tokenizer helpers to follow project conventio

b9286

B9286 introduces support for HybridDNATokenizer as its own vocabulary type (LLAMA_VOCAB_TYPE_HYBRIDDNA), enabling proper tokenization of DNA sequences by routing raw text through a DNA-aware splitter

b9292

Release b9292 of llama.cpp fixes a memory leak in the server context where speculative decoder, draft context, and draft model were not properly freed during destroy(), causing VRAM leaks on sleep/res

Boundary-targeted Membership Inference Attacks on Safety Classifiers

arXiv:2605.22373v1 Announce Type: cross Abstract: Safety classifiers are essential safeguards within generative AI systems, filtering harmful content or identifying at-risk users when interacting with

Broadening Access to Transportation Safety Data with Generative AI: A Schema-Grounded Framework for Spatial Natural Language Queries

arXiv:2605.21712v1 Announce Type: new Abstract: Transportation safety analysis requires integrating crash records, roadway attributes, and geospatial data through GIS-based workflows, but access remai

FastTab: A Fast Table Recognizer with a Tiny Recursive Module and 1D Transformers

arXiv:2605.22422v1 Announce Type: new Abstract: Table structure recognition (TSR) requires both table-level coherence (row/column counts, headers, spanning cells) and precise separator localization. W

GALAR-TemporalNet v2: Anatomy-Guided Dual-Branch Temporal Classification with Bidirectional Mamba and Dual-Graph GCN for Video Capsule Endoscopy -- after competition results

arXiv:2605.22209v1 Announce Type: new Abstract: Video Capsule Endoscopy (VCE) poses a challenging multi-label temporal classification problem, requiring simultaneous localization of 8 anatomical regio

Heartbeat-Bound Hierarchical Credentials: Cryptographic Revocation for AI Agent Swarms

arXiv:2605.20704v1 Announce Type: cross Abstract: Autonomous AI agents that spawn sub-agent swarms create a safety gap: existing credential revocation mechanisms, OAuth~2.0 introspection, OCSP, and W3

local-aiarxiv-cs-ai

Hypergraph as Language

arXiv:2605.21858v1 Announce Type: new Abstract: Large language models (LLMs) have recently shown strong potential in modeling relational structures. However, existing approaches remain fundamentally g

I built a free demo for Pixal3D (Tencent new image-to-3D model)

Pixal3D is a Tencent image-to-3D model that generates high-fidelity 3D assets from a single image by explicitly lifting pixel features into 3D through back-projection to establish direct pixel-to-3D c

img2vid: ComfyUI doesn't find the spatial upscaler

This Reddit post discusses a common ComfyUI issue where users cannot locate spatial upscaler models for img2vid workflows. The problem typically stems from placing spatial upscaler models in the wrong

Intelligence per Watt: Measuring Intelligence Efficiency of Local AI

arXiv:2511.07885v4 Announce Type: replace-cross Abstract: Large language model (LLM) queries are predominantly processed by frontier models in centralized cloud infrastructure. Demand growth strains t

Look-Closer-Then-Diagnose: Confidence-Aware Ultrasound VQA via Active Zooming

arXiv:2605.21652v1 Announce Type: new Abstract: Vision-Language Models (VLMs) have significantly advanced medical visual question answering, yet their performance in ultrasound remains suboptimal. In

MTP is available in LM Studio 0.4.14. Sound on.

LM Studio version 0.4.14 introduced support for MTP (likely Model Transfer Protocol or a similar feature), with audio functionality enabled. This update was announced on the official LM Studio X/Twitt

local-ailm-studio--x

Runaway token costs and sovereignty concerns are driving enterprises back to the desktop

The AI PC is being fundamentally redefined as agentic workloads push the boundaries of what local compute can deliver — and as runaway cloud token costs force enterprises to rethink where inference ac

local-aisiliconangle

SceneAligner: 3D-Grounded Floorplan Localization in the Wild

arXiv:2605.22581v1 Announce Type: new Abstract: Many public buildings provide floorplans with a 'you are here' indicator to help visitors orient themselves. Floorplan localization seeks to computation

SegCompass: Exploring Interpretable Alignment with Sparse Autoencoders for Enhanced Reasoning Segmentation

arXiv:2605.22658v1 Announce Type: new Abstract: While large language models provide strong compositional reasoning, existing reasoning segmentation pipelines fail to transparently connect this reasoni

Structured-Sparse Attention for Entity Tracking with Subquadratic Sequence Complexity

arXiv:2605.22476v1 Announce Type: cross Abstract: Entity tracking requires maintaining and updating latent states for entities and attributes over long sequences. Recent task-specific attention operat

This is what we love to see. Welcome to the community 🤝

ComfyUI posted a welcoming message to new community members on X (Twitter), using a celebratory tone to encourage engagement and belonging within their user base. The post reflects community-focused v

local-aicomfyui--x

Training low resolution then switching to high resolution later?

A discussion of progressive training approaches for diffusion models where training begins at lower scale factors and progressively increases to target scale factors, leveraging previously trained mod

using ltx 2.3 i2v 3d animation with reference voice using TalkVid Lora.

ID-LoRA (Identity-Driven In-Context LoRA) enables identity-preserving audio-video generation by simultaneously dictating the scene's visual content, environmental acoustics, and speaking style while p

VBFDD-Agent for Electric Vehicle Battery Fault Detection and Diagnosis: Descriptive Text Modeling of Battery Digital Signals

arXiv:2605.20742v1 Announce Type: new Abstract: With the rapid proliferation of electric vehicles, the safety and reliability of lithium-ion batteries have become critical concerns. Effective anomaly

local-aiarxiv-cs-ai

VRXU-net: A Deep Learning Approach for Brain Ischemic Stroke Lesion Detection and Segmentation in T1W MRI

arXiv:2605.21633v1 Announce Type: cross Abstract: When the blood supply to the brain is obstructed by a clot, oxygen delivery to brain tissues becomes insufficient, leading to cellular necrosis. In he

Which Way Did It Move? Diagnosing and Overcoming Directional Motion Blindness in Video-LLMs

arXiv:2605.22823v1 Announce Type: new Abstract: Video Large Language Models (Video-LLMs) have made rapid progress on temporal video understanding, yet many fail at a basic perceptual primitive: signed

AceStep 1.5 - lora trained on first two albums of Modern Talking band

AceStep 1.5 is a LoRA (Low-Rank Adaptation) model trained on the first two albums of the Modern Talking band, designed for use with Stable Diffusion image generation. This LoRA likely enables the gene

AIGaitor: Privacy-preserving and cloud-free motion analysis for everyone, using edge computing

arXiv:2605.21421v1 Announce Type: new Abstract: Motion capture is the gold standard for measuring human movement, but clinical use remains limited by cost, technical complexity, and privacy concerns.

b9255

Release b9255 of llama.cpp features a Hexagon HMX quantized matmul rework (#23368), including updates to debug logging, dequantization logic using HVX vectors, removal of non-pipelined quantization op

b9257

llama.cpp release b9257 optimizes operations in the Vulkan IM2COL shader and improves code formatting . The release includes prebuilt binaries for multiple platforms including macOS, Linux, Android, a

b9259

Release b9259 of llama.cpp fixes a nullptr crash in the speculative common module's get_devices_str function (PR #23386). The fix addresses an issue where ggml_backend_dev_by_name appends a nullptr se

b9260

Release b9260 of llama.cpp includes OpenCL backend refactoring that improves initialization, GPU identification, and performance by caching global memory size in device context. llama.cpp enables LLM

b9263

Release b9263 of llama.cpp includes a merge of HunyuanOCR into HunyuanVL with fixes to OCR vision precision. The update consolidates OCR functionality into the HunyuanVL projector while maintaining co

b9264

b9264 is a llama.cpp release that includes improvements to HunyuanVL model support, merging HunyuanOCR functionality and fixing vision precision issues. This build represents an intermediate developme

b9266

b9266 is a release tag for llama.cpp, a project for LLM inference in C/C++. This release represents a specific build version of the llama.cpp framework, which enables LLM inference with minimal setup

b9270

Release b9270 adds support for the HybridDNATokenizer used by the Carbon-3B model family, implementing a new BPE pre-type for tokenizing DNA sequences. The tokenizer handles DNA k-mers with fixed 6-me

b9273

Release b9273 of llama.cpp introduces support for the HybridDNATokenizer with new pre-type and dispatched tokenization logic, alongside pure helper functions for DNA k-mer processing and conversion ut

b9275

Release b9275 of llama.cpp includes optimization of the Metal concat kernel and fixes to the GGML_OP_SET kernel threads . The release extends test coverage for copy operations with different source an

Closed-form predictive coding via hierarchical Gaussian filters

arXiv:2605.20293v1 Announce Type: new Abstract: Predictive coding (PC) offers a local and biologically grounded alternative to backpropagation in the training of artificial neural networks, yet to dat

local-aiarxiv-cs-lg

Conformal Selective Acting: Anytime-Valid Risk Control for RLVR-Trained LLMs

arXiv:2605.20270v1 Announce Type: new Abstract: A local specialist LLM, fine-tuned with reinforcement learning from verifiable rewards (RLVR) on operator-local data, is installed in a regulated organi

local-aiarxiv-cs-lg

Control FLUX.2 with reference images instead of training a LoRA — demo

This post demonstrates using reference images with control maps in FLUX.2 instead of training LoRAs, allowing users to fuse a reference image with pose or edge control data while preserving identity a

Creé mi propio agente de IA local centrado en Ollama y la ejecución de tareas prácticas.

This post describes a user's experience building a local AI agent using Ollama, an open-source tool for running large language models locally. The article likely covers the practical implementation st

local-air-ollama

Diffusion Models Memorize in Training -- and Generalize in Inference

arXiv:2603.13419v2 Announce Type: replace Abstract: Diffusion models generalize well in practice. However, an optimal diffusion model fully memorizes the training data and therefore fails to generaliz

local-aiarxiv-cs-lg

DNACHUNKER: Learnable Tokenization for DNA Language Models

arXiv:2601.03019v4 Announce Type: replace-cross Abstract: DNA language models are increasingly used to represent genomic sequence, yet their effectiveness depends critically on how raw nucleotides are

E-ReCON: An Energy- and Resource-Efficient Precision-Configurable Sparse nvCIM Macro for Conventional and Spiking Neural Edge Inference

arXiv:2605.20717v1 Announce Type: cross Abstract: This work presents E-ReCON, a 16 Kb energy and resource-efficient digital compute-in-memory (DCIM) macro based on a compact 3T1R ReRAM bitcell for edg

End-to-End Unmixing with Material Prompts for Hyperspectral Object Tracking

arXiv:2605.20569v1 Announce Type: new Abstract: Hyperspectral imagery encodes rich material properties that can improve tracking robustness under appearance ambiguity, illumination change, and backgro