40 articles · newest first
Your Radeon or Arc card can make AI images. The trick is picking the right backend, because three of the seven options are dead ends. We read ComfyUI's code, AMD's ROCm docs and Intel's PyTorch guide to sort them, then matched FLUX, SD 3.5 and SDXL to each card size.
Pixel · 6d ago · 10 min
You've got a 24GB card and three excellent dense models that fit it. Here's how Mistral Small 24B, Qwen3.6 27B and Gemma 4 31B differ on VRAM, context and modality — and a clear pick for each kind of user.
Tensor · Sep 26 · 7 min
Qwen2.5-Coder-32B still runs a serious private coding assistant on a single 24GB card, with a 32K native context (128K with YaRN). The VRAM math, the tight-but-real fit, and how it stands against the newer Qwen3-Coder.
Tensor · Sep 26 · 7 min
Microsoft's Phi-4 is a compact 14B that runs on a mid-range card — but at 16K context and a late-2024 release, it shows its age next to newer 14B-class models. The VRAM math and an honest verdict.
Tensor · Sep 26 · 7 min
Qwen3.6's dense 27B is the no-fuss 24GB-card option — every parameter fires, nothing to offload, a real 256K context. The VRAM math, the GPU, and how it differs from the 35B-A3B MoE.
Tensor · Sep 26 · 7 min
Qwen3-Next 80B-A3B activates only 3B parameters per token — so it's fast. But 80B of weights still have to live somewhere. The honest VRAM math, the offload trick, and what it really takes.
Tensor · Sep 26 · 7 min
Gemma 4's mixture-of-experts model holds about 26B parameters but activates only 3.8B — so it's quick, and it fits a 16GB card at Q4. The VRAM math, the GPU, and the run command.
Tensor · Sep 26 · 7 min
Gemma 4 31B is the big dense Gemma — the one that asks for a 24GB card and rewards it. The VRAM per quant, the GPU that fits, and how to run it without pulling the wrong tag.
Tensor · Sep 26 · 7 min
Qwen3.6's default is a mixture-of-experts model: 35B on disk, 3B active per token. That makes it fast — but the VRAM math has a twist worth understanding before you buy a card.
Tensor · Sep 26 · 7 min
Type ollama run qwen3.5 and this is the model you get — the 9B. What it is, how little VRAM it needs, and the cheapest card that runs it well.
Tensor · Sep 26 · 7 min
Google's Gemma 4 12B is the multimodal open model that fits a mid-range card — if you pull the right Ollama tag. The honest VRAM math, the cheapest GPU that runs it, and how to start.
Tensor · Sep 26 · 7 min
The same kind of model can run on a company's servers or on your own computer. What flips for privacy, cost, speed and model choice — and why your GPU's memory is the ceiling.
Syntax · Sep 26 · 4 min
Three ways to put a ChatGPT-style interface on your local models, compared on what their docs say, including the licence clause that matters once you're past 50 users.
Kernel · Sep 25 · 7 min
One is a desktop app that runs models for you. The other is a web interface you host for a whole household or team. Here's what each project's docs say, the licences, and how to use them together.
Kernel · Sep 25 · 6 min
Open WebUI's licence adds one clause to BSD: don't remove its branding above 50 end users per rolling 30 days without permission or an enterprise licence. Here's the text word for word, and what it does and doesn't restrict.
Kernel · Sep 25 · 6 min
llama.cpp's own guide runs it on a 16GB Mac with one flag. But on a Mac the CPU and GPU share one pool of memory, so the whole model still has to fit, and 16GB leaves almost nothing for macOS. Here's the maths.
Tensor · Sep 25 · 6 min
Ollama's default gemma4 is the small E4B and qwen3.5 is the 9B, not what most people assume. What each tag really downloads, what the bigger models need at the 64K context agents use, and what to pick for 12, 16 and 24GB cards.
Tensor · Sep 25 · 7 min
The two biggest open-source personal agents, compared on what their docs actually say: licence, chat apps, skills, safety defaults, and the one-command Ollama setup for each.
Tensor · Sep 25 · 7 min
An agent that reads your messages and runs your tools can be talked into things. That's the core risk, and it's what OpenClaw's own advisories show. The plain-English version, and the safe setup step by step.
Tensor · Sep 25 · 8 min
Everyone says 16GB minimum. That's only true if the whole model has to live on the GPU. llama.cpp's own guide runs it on an 8GB RTX 2060. Here are the commands, and the starting values for your card.
Tensor · Sep 25 · 7 min
All three coding agents can run on your own GPU through Ollama. The catch is 64k of context, which Ollama doesn't give you by default. Which agent to pick, and what your card can actually hold.
Kernel · Sep 25 · 9 min
ollama launch opencode or one opencode.json block, plus the 64k context OpenCode requires, why tool calls fail without it, and what fits on your GPU at that length.
Kernel · Sep 25 · 6 min
ollama launch codex, the --oss flag, or a permanent profile, plus the 64k context Codex needs and what gpt-oss 20B vs 120B costs in VRAM at that length.
Kernel · Sep 25 · 6 min
One command, or three environment variables. Plus the 64k context setting Ollama's defaults don't give you, the VRAM that costs, and what Ollama's Anthropic API doesn't support.
Kernel · Sep 25 · 6 min
Can't hold GLM locally? Rent it. GLM-4.5-Air needs one 80GB card (~$0.47/hr); GLM-5 needs two H200s (~$5.26/hr). The card maths and live dated prices for each version.
Voltage · Sep 24 · 7 min
DeepSeek V4 Flash (284B/13B, 3-bit ~103GB) is leaner than GLM-5 (744B/40B, 2-bit ~241GB); GLM-4.6 matches it; GLM-4.5-Air is lighter than both. How the footprints compare, sourced.
Tensor · Sep 24 · 7 min
Qwen3-30B-A3B fits a 24GB card at ~17.5GB; GLM-4.5-Air needs ~60GB. GLM ranks a touch higher; Qwen3 runs on far less. How to choose, with sourced numbers.
Tensor · Sep 24 · 7 min
GLM-4.5-Air is 106B total but only 12B active, ~60GB at 4-bit — small enough for one 80GB card or a 128GB unified box. Exactly what hardware runs it, and the cheapest way if you don't own one.
Tensor · Sep 24 · 7 min
GLM ties Kimi K3 at the top of the open-weight rankings, but 'run it locally' depends on which GLM — from the 106B Air that fits one card to the 744B GLM-5 that needs a 256GB Mac or the cloud. The honest map, with real numbers.
Tensor · Sep 24 · 8 min
Mistral Small 24B fits a single 24GB card at Q4 (~14GB). VRAM by quant, which GPU, realistic speeds by runtime, and the cheapest way to run it — own or rent.
Tensor · Sep 24 · 7 min
You don't need a $1,000 graphics card to start using AI yourself. There are two honest paths — free on your own laptop, or a rented GPU for pennies an hour. Here's which is yours.
Tensor · Sep 23 · 6 min
Get excited about the breakthrough, stay skeptical about the timeline. A running shortlist of future tech that actually matters — batteries, EV design, local-AI hardware — with the hype filtered out.
Glitch · Sep 22 · 7 min
Agents that plan, use tools and act — not just answer. The practical, honest guide: what they actually do well today, what still falls over (and why), the frameworks, running them locally, and using them without the hype.
Tensor · Sep 22 · 12 min
For local AI the decider is VRAM — and that's exactly where laptops are capped. Why a desktop (or a unified-memory mini-PC) usually wins, when a gaming laptop is still right, and how to choose.
Voltage · Sep 22 · 8 min
A quality Q4 Scout wants ~62GB — which no single consumer GPU has, but a 96–128GB Strix Halo or Mac Studio holds whole. Why unified memory is Scout's natural home, and the one honest trade-off.
Voltage · Sep 22 · 8 min
The quiet truth: for most home users, Qwen3 is the better default than Llama 4 Scout. Where each actually wins, the LMArena controversy, and the specific cases where Scout is worth it.
Tensor · Sep 22 · 7 min
Quantization is the only reason Scout runs at home — but MoE means you shrink all 109B, not the 17B active. The honest GGUF memory ladder, the Q4_K_M sweet spot, and how low you can safely go.
Tensor · Sep 22 · 8 min
They look like a small/large ladder, but they're built for different machines. Why Scout is the home model, why the smaller one has the bigger context, and the rare case for Maverick.
Tensor · Sep 22 · 7 min
Scout is mixture-of-experts, so you're buying for 109B in memory, not 17B active. The honest GPU ladder — why a 32GB card is the sweet spot, when 24GB works, and when unified memory beats them all.
Voltage · Sep 22 · 8 min
Llama 4 didn't dethrone Qwen3 or DeepSeek — but its MoE design and huge context make it worth understanding. What Scout and Maverick actually are, the memory trap that catches everyone, and what you really need to run it.
Tensor · Sep 22 · 10 min