Hardware Corner @hwcorner

We just published a new article on running Gemma 4 locally, based on our own testing. We looked at the 26B and 31B models from a practical hardware perspective, focusing on VRAM usage, context scaling, and real performance on consumer GPUs. hardware-corner.net/hardware-for-g…

New article is up. We took a close look at the Intel Arc B70 32GB and its first local LLM benchmarks, including a direct comparison with the RTX 3090. hardware-corner.net/intel-arc-b70-…

Introducing Unsloth Studio ✨ A new open-source web UI to train and run LLMs. • Run models locally on Mac, Windows, Linux • Train 500+ models 2x faster with 70% less VRAM • Supports GGUF, vision, audio, embedding models • Auto-create datasets from PDF, CSV, DOCX • Self-healing tool calling and code execution • Compare models side by side + export to GGUF GitHub: github.com/unslothai/unsl… Blog and Guide: unsloth.ai/docs/new/studio Available now on Hugging Face, NVIDIA, Docker and Colab.

Just published a new article on the first Apple M5 Max local LLM benchmarks. We compared the MLX results running 120B models on a 128GB MacBook Pro with our RTX Pro 6000 and RTX 5090 tests. Interesting results for anyone running large models locally. hardware-corner.net/m5-max-local-l…

If you run local LLMs, this deal is worth a look. A Lenovo ThinkStation P3 Tower Gen 2 with an RTX Pro 5000 Blackwell 48GB is listed for about $4,719, which is close to the normal price of the GPU alone. hardware-corner.net/local-llm-48gb…

Interesting shift in the OpenClaw ecosystem. PicoClaw in Go and ZeroClaw in Rust just proved the agent layer can run on $10 Raspberry Pi class hardware. No need for a Mac mini just to orchestrate tools. hardware-corner.net/openclaw-whith…

We just published a short article on what hardware you need to run the new Qwen3 Coder Next 80B A3B locally, including VRAM requirements and speed benchmarks. hardware-corner.net/qwen3-coder-ne…

DeepSeek V4, Qwen 3.5, and GLM 5 are expected soon, and they could matter a lot for people running LLMs locally. hardware-corner.net/deepseek-v4-qw…

Just published a deep dive on what it actually takes to run ClawdBot locally. It breaks down which hardware makes sense at different model sizes, where unified memory works, where GPUs still win, and why agent workflows are much harder than chat. hardware-corner.net/best-computers…

Just saw a new Strix Halo mini PC from Sixunited, and the size is what really stands out. At only 1.2 liters, it’s the smallest Ryzen AI Max+ 395 system so far. hardware-corner.net/1-2l-strix-hal…

Early listings hint that NVIDIA N1X may be coming to LLM-capable laptops. Unified memory and high bandwidth could finally make serious local inference possible on a single portable system. Still early, but worth watching. hardware-corner.net/llm-laptops-nv…

The RTX 5060 Ti 16GB has quietly moved up to around $550+ at most stores. If you’re building for local LLMs, there’s still one retailer selling close to MSRP, which is getting rare. hardware-corner.net/rtx-5060-ti-16…

GPU prices are moving up again. High-VRAM cards are getting harder to find, and there’s no near-term GPU refresh to ease the pressure. If you’re building a local LLM box, this may be a window worth paying attention to. hardware-corner.net/gpu-prices-ris…

DeepSeek’s Engram research takes a fresh look at why LLMs consume so much VRAM, and whether separating reasoning from memory could make large models more accessible on local hardware. hardware-corner.net/deepseek-ngram…

One of the most practical GPUs for running local LLMs, the RTX 5060 Ti 16GB, may be entering a supply squeeze. Prices are already rising in Europe, while the US market is still mostly stable. hardware-corner.net/rtx-5060-ti-16…

For anyone running LLMs locally, a 16GB GPU at a reasonable price is getting harder to find. This card can handle mid-size models and long context well, and current pricing still makes sense compared to other GPUs. hardware-corner.net/rtx-5060-ti-16…

PewDiePie built a $40K workstation with 10 GPUs and 424GB of VRAM — enough to run 600B+ LLMs locally. Built on a Threadripper Pro platform with undervolted 48GB RTX 4090s. hardware-corner.net/pewdiepie-ai-p…

The first DGX Spark benchmarks are finally here. It’s fast, but where it wins (and where it falls short) might surprise you. Check out how it stacks up in real LLM tests. hardware-corner.net/first-dgx-spar…

For anyone hitting the VRAM wall with local LLMs: there’s a 96GB VRAM card from Huawei showing up around $1500. Not really a gaming GPU, but an interesting option for inference experiments. hardware-corner.net/huawei-atlas-3…

Just tested GPT-OSS 120B with MoE offloading in llama.cpp. Offloading expert layers to CPU makes a difference on both RTX 3090 and 5090. hardware-corner.net/guides/gpt-oss…