
Qwen3.8-2.4T-A95B: the math of running 2.4T open-weight parameters, 95B active
The first Qwen-Max-class open-weight model weighs ~1.3 TB at Q4: no single consumer node fits it. The math on amortization, power, and speed for multi-node.
Your AI is always ready: like your own car versus a taxi. No queues, no request limits.
Your data never leaves your hardware. Period.
No subscription, no per-request cost. Once you have the hardware, the AI is yours.
The KIWI paper (arXiv) flips the device-initiated communication hierarchy: reverse offload on host threads, exposed as a programmable interface, beats GPUDirect backends — up to 2.3x lower expert-parallel MoE communication latency on AMD GPUs and -29% time per token in vLLM.
The WAM-Policy paper teaches arm-less quadrupeds to walk on three legs while manipulating with the fourth, following language instructions: on-policy distillation from world action models lets the policy follow commands, and the project is open-sourced.
An independent benchmark measures 22 local LLMs (the specific quantizations that run on a gaming PC) with agentic coding challenges in isolated containers — ~10 hours per full run — and compares Qwen 3.6 fine-tunes against alternative dense models like Qwen 3.8 Flash Next.
NVIDIA releases Riva-Translate-4B-Instruct v1.1: a 4B-parameter translation model (BF16, ~8.4 GB) distilled from Mistral-NeMo-Minitron-8B-Base, covering 12 languages and dialects with performance comparable to EuroLLM-9B-Instruct on FLORES and WMT24.

The first Qwen-Max-class open-weight model weighs ~1.3 TB at Q4: no single consumer node fits it. The math on amortization, power, and speed for multi-node.

Every conversation stored word for word in a SQLite file, with no model call at all: search runs on local embeddings and retrieval costs 3.1k tokens. 91.8% on LongMemEval, zero tokens to save.

64.8B parameters, Apache 2.0, 1M of context and an NVFP4 checkpoint designed for 128 GB of unified memory: the math and the criteria for running a frontier dense locally.
DeepSeek's MoE model with 552B total parameters and 8-16B active per token. 1M-token context, MIT license: free commercial use. Built for online serving, too large for local use on a single PC.
Terminal-free desktop app for downloading and running GGUF models: GUI, built-in model search and a local server. The simplest way to get started.
Bespoke Labs' 9B decision model: it runs entirely locally via Ollama 0.35 and answers typed questions (true/false, numeric scale, multiple choice) through the new systemone API. Built for fast classification in automated pipelines, no cloud.
Frontier-class open model with 131K-token context. Needs 48-80GB VRAM, a Mac with unified memory, or a multi-GPU setup.
Meta's largest open model: closed-frontier quality but datacenter requirements. In practice it needs high-VRAM multi-GPU or the cloud.
A free, step-by-step PDF: hardware, models, tools and first steps to bring AI into your business — no cloud, no data sharing.
Download the free guideFree. No strings attached. No email address required.
The newsletter is free and only needs your email. With a full account, the site works for you: save your setups, personalize your profile, and keep everything in control.
Save your hardware and model plans: find them all, ready, when you come back.
Name, bio and preferences: your space on the site, yours forever.
Easy access to the full newsletter archive and the weekly selection of what's changing in local AI.
Manage email, password and subscription whenever you like. No commitments, no spam.