Boosting LLM Reasoning Without Adding Parameters: The Principle of 'Virtual Logic Depth'
Google DeepMind has introduced 'Virtual Logic Depth,' a scaling dimension that boosts LLM reasoning capabilities without increasing parameter size by looping da
#LocalLLM
TAU-HOME.COM stories tagged LocalLLM.
Magnitude Open-Source LLM Inference Engine Released: Auto-Tuning Kernels Deliver Over 2x Faster Decode Than Ollama
Tom Greenwald has announced Magnitude, a 100% free and open-source local LLM inference engine that auto-tunes compute kernels to user hardware, averaging 2.19x
Humanizer: Open-Source 12B Local Model That Rewrites Mechanical AI Drafts into Human Tone
Ranked #2 on Hugging Face trending, 'Humanizer' is an open-source 12B local model designed to rewrite AI-generated drafts—such as emails, essays, and reports—in
sparkDash v2.0 Released: Redesigned Local LLM Fleet Dashboard with In-Card Launches, Benchmarks, and Expanded Monitoring
Local LLM fleet dashboard sparkDash v2.0 is out with a redesigned UI, direct model launches from cards, Decode/Prefill/Quality/Tool Eval benchmark pages, and ne
Ghost AI Unveils 'Core', an On-Device Personal AI Computer with RTX PRO 4000 Blackwell
Ghost AI has unveiled Core, a personal AI computer running local models 24/7 on-device. Powered by NVIDIA RTX PRO 4000 Blackwell SFF, it hits 87 tok/s on Qwen-3
Qwen3.8-Flash-Next NVFP4 Benchmarked on Single DGX Spark: 1M Context with FP8 KV Cache
A benchmark of Qwen3.8-Flash-Next NVFP4 on a single 128GB DGX Spark demonstrates 1M context with FP8 KV cache, 4-stream concurrency, and full multimodal input s
Aleph Alpha 78B MoE Kolibri-1 Community GGUF Released: Reaching 13-15 TPS on CPU
Community GGUF quantizations for Aleph Alpha's 78B MoE model Kolibri-1 are now available, achieving 13-15 TPS on a Ryzen 7 7800X3D CPU using 128GB RAM and patch
Local AI Runtime Benchmarks: Comparing exllamav3, llama.cpp, vLLM, and TensorFold Across 16GB to 256GB Memory Classes
Empirical benchmark results comparing local LLM inference speeds (tok/s) and Time to First Token (TTFT) across RTX 5060 Ti (16GB), RTX 5090 (32GB), and GB10 (12
Strata v0.1.34 Released: AMD GPU Acceleration on Windows and Built-in MCP Server Support
Open-source local LLM engine Strata v0.1.34 adds official AMD GPU acceleration on Windows and a built-in MCP server for direct AI coding agent integration.
Compressing Base Context from 30k to 10k for Local LLM Prefill: Coding Harness 18.3.0 Optimization
Coding agent harness 18.3.0 slashes base prompt token usage from ~30k to 10k for faster local model prefill, consolidates tools into read/write/bash, enables de
Magnitude: Open-Source Local LLM Profiling, Hardware Tuning, and Agent Harness Connector
Magnitude profiles your PC hardware, ranks local LLMs with estimated tok/s, applies speculative decoding, and bridges local models to Pi, OpenCode, and Claude C
Qwen3.8-27B Local Agent Coding Tips: Why 2.50 bpw GSQ-RCO IQ2_XS Beats Raw Benchmark Scores
Static 98.2% retention benchmarks miss agentic coding breakdown. How switching to ISTA-DASLab GSQ-RCO IQ2_XS (2.50 bpw) on a 12GB RTX 3080 Ti restores task reli
Bonsai 2 27B: 5.95GB Ternary Model Gets Weight-Level Refusal Abliteration CRACK Build
The dealignai team has released a weight-level abliterated CRACK build of PrismML's 5.95GB ternary Qwen3.8-27B model Bonsai 2, eliminating refusal behavior with
Run All-in-One AI Locally Without a GPU: Open-Source Runtime 'LocalAI'
LocalAI runs LLMs, voice, vision, and agents locally without cloud API fees or data leakage. Explore OpenAI-compatible endpoints, pluggable backends, and CPU op
Uncensored Nex-N2.5-mini GGUF: Running a 35B MoE Locally on an 18GB VRAM Card
An uncensored GGUF build of Nex-N2.5-mini (IQ4_XS) enables local execution on an 18GB GPU. Explore its 35B MoE architecture, 262K context, vision mmproj, and ll
One-Click Kit Runs Qwen3.8-27B EXL3 on 16GB NVIDIA GPUs via OpenAI-Compatible API
MiaAI-Lab has released an open-source one-click serving kit that sets up Qwen3.8-27B using turboderp's EXL3 quantization on a single 16GB VRAM NVIDIA GPU with l