AbstractPhil released Beatrix V3, a 376-million-parameter byte-level language model with a 256-token vocabulary and 4,096-byte context, trained on 64.4 billion bytes across two RTX 5090 cards in 295 h…
A 27B NVFP4-quantized model served by vLLM on a Windows 11 desktop with an RTX 5090 (32 GB VRAM) dropped from roughly 77 tokens per second to about 7 tokens per second during Microsoft Teams video cal…
FastH3 V2, an eight-step video-and-audio generation model whose weights previously required 138 GiB and data-center GPUs, now runs on a single consumer machine — an RTX 5090, RTX 4090, RTX PRO 6000, D…
Chutes AI trained an 8 billion parameter model on 240 RTX 5090 consumer GPUs spread across 30 hosts in 13 countries for approximately $6,500, or about $11 per billion tokens processed, Jon Durbin said…
Signalnine released PotemkinOS, an MIT-licensed Linux image with no userland — no /bin, shell, coreutils or package manager — in which a local model writes the userland from a prompt using eight tools…
A personal LLM user reports that Qwen3.8-27B running locally via llama.cpp on an RTX 5090 laptop GPU with 24 GB of VRAM now handles Q&A, coding, sysadmin and web research without a cloud provider, del…
Prism ML released Bonsai 2 27B, a ternary-quantized version of Qwen3.8-27B that stores weights as -1, 0, or +1 and fits in roughly 6 to 8.6GB while retaining 98.2% of the original FP16 model's benchma…
ComfyUI v0.36.0 ships optimizations that make the MiniMax H3 video VAE encode up to ~2.2x faster and decode ~1.4-2.7x faster, cutting a 1344x768, 129-frame encode-and-decode round trip from 24.3 to 12…
A developer published a llama-server configuration that runs a Qwen 3.8 27B NVFP4 model with MTP speculative decoding on an RTX 5090, reporting throughput of 90-120 tokens per second. The setup uses a…
NVIDIA listed the RTX PRO 5500 Blackwell, an 84GB GDDR7 ECC workstation card with 21,760 CUDA cores, 1,398 GB/s of bandwidth, 600W power draw and PCIe Gen 5, as "Coming Soon" on nvidia.com in Septembe…
A hobbyist software developer running three machines — a 9800X3D with an RTX 5090, a 5900X with an RTX 4090, and a 3900 with an unused RTX 2080ti — is asking for setup advice on running local LLMs acr…
Kaarel Kaarelson released LingBot-World 2.0, a 1.3B-parameter world model that runs at 16.1 FPS on a single RTX 5090, 2.7× faster than the original paper's code at 6.0 FPS with no loss of output quali…
Researchers submitted SSD-LLaMA to arXiv on 16 Sep 2026, an SSD-native local Mixture-of-Experts inference system that runs trillion-parameter models at over 1 token/s on a single RTX 5090 with no more…
A developer built a local personal AI assistant running on two 12 GB consumer GPUs, combining an LLM with RAG, graph memory, MCP tools, and voice input/output to handle documents, email, calendar, Tel…
NVIDIA released Personal AI Router (PAIR) in beta, a tool that distributes individual AI inference requests across multiple computers on a local network, integrating with Ollama and LM Studio without …
A developer published a serving configuration for the turboderp/Qwen3.8-Flash-Next-exl3-4.05bpw EXL3 quant, running the large MoE model split across an RTX 5090 and RTX 4090 with 192 GB of DDR5 system…
Thesys Dev released OUI-1, the world's first model for Generative UI, a finetuned DiffusionGemma model that writes user interfaces in openui-lang and runs on consumer-grade GPUs like the RTX 5090 at F…
Nvidia unveiled Personal AI Router (PAIR), a free open-source tool that distributes local AI workloads across available computers on a network, at IFA 2026. PAIR supports Windows, macOS, and Linux and…
A new version of the DLSS Unlocked mod on NexusMods enables DLSS 5 and multi-frame generation on Nvidia RTX 20, 30, and 40-series GPUs, according to VideoCardz. The mod combines Nvidia's neural render…
A developer built a neural block texture compression system that runs entirely on the GPU using CUDA 13.1, combining per-pixel selectors, per-4×4-block latents, and a tiny 1,767-weight MLP decoder tra…