Featured
Local LLM
New 2026
Use local Qwen3.8-27B and official MiniMax-H3 Skills in ComfyUI to generate H3 prompts. Runs a standalone local llama-server , inference operates fully offline with no internet access.
New 2026
LM Studio Vulkan RAM patch for the ASUS ROG Flow Z13 128GB only. Stops llama-server from pinning a 25GB host copy into the 32GB Windows-visible RAM window.
New 2026
Qwen3.8-27B on Apple silicon, end to end: conversion integrity (vision tower + MTP head preserved), a split-K Metal kernel, two speculative paths and a reversed verdict, bitwise two-box TB5 prefill - with the full ledger, harness, and raw measurement receipts.
New 2026
Fast Qwen3.8-27B inference on one RTX 3090: ReplaySSM, MTP3, reasoning effort, C1-C8 batching, and native Windows and Linux builds.
New 2026
A first class Hermes model provider plugin for llama.cpp and llama-swap.
New 2026
Local Chrome extension: score the job tab you're already on. Ollama, no crawl, no auto-apply.
New 2026
Maestro AI port to AMD GPU
New 2026
Official desktop client for GLM-5.3. 1M-token codebase auditing, RL-powered zero-day discovery, and exploit chaining. FREE unmetered access until October!
New 2026
Qwen3.8-27B at 34-38 tok/s on DGX Spark (GB10) — one-command SGLang + NVFP4 + DSpark setup, systemd, Claude Code ready
New 2026
Qwen3.8-27B on a single RTX 3090: 416 tok/s batched or 25 ms/token single-user, 150k context, vLLM
New 2026
Qwen3.8-27B-NVFP4 on one DGX Spark — measured MTP/DFlash/DSpark, recipes, raw logs
New 2026
Serving Qwen3.8-27B (Unsloth UD-Q4_K_XL GGUF) at 23 tok/s over a 104,192-token context on a single NVIDIA L4 24GB, llama.cpp + MTP speculative decoding. One-command GCP provisioning, tuning sweep, quality verification, and a binary-searched context ceiling.
Other topics
Browse other topics on Trendshift
A
N