Skip to content
Independent AI tools and industry intelligence

Find the AI tools your hardware can actually run.

Search a hardware-aware catalogue, build a compatible local stack, and follow source-led AI news without paid placements or invented benchmarks.

Build your stack →Read the latest AI brief
Featured workflow engineComfyUIThe nodal workflow engine for serious diffusion.Open sourceLocal workflowHardware mappedOpen datasheet →
100+
tools tracked
11
use-case categories
0
paid placements
Latest AI brief · July 13, 2026

AI Hub Search Weekly: US Safety Policy, OpenAI Red Teaming, and Developer Releases

OpenAI advanced AI safety via state-federal governance and GPT-Red self-play red teaming. Hugging Face covered agent building, model routing, and Inkling. NVIDIA posted DeepStream, USD, and CUDA guides. Ollama and llama.cpp shipped runtime builds.

  1. 01OpenAI outlines reverse federalism AI governance approach
  2. 02OpenAI releases GPT-Red automated red teaming system
  3. 03Hugging Face publishes Shippy agent-building lessons
Read all 7 stories →
Documentation-led research

Decisions, not another tool list

Manually authored guides that show the upstream evidence, the assumptions, and the point where a hardware claim becomes workload-dependent.

All research guides →
Guide 01 · 12 min

ComfyUI VRAM planning

There is no single ComfyUI minimum. Buy for the largest model, resolution, and sequence you expect to use repeatedly—not for the smallest workflow you can make load once.

Read the evidence →
Guide 02 · 14 min

LoRA training hardware

Training memory depends on what stays trainable and resident, not only checkpoint size. SD 1.5 is the forgiving starting point; SDXL raises the baseline; Flux examples often target 24 GB and should not be generalized to every configuration.

Read the evidence →
Guide 03 · 13 min

Local LLM memory planning

Model-file size is a lower bound, not a complete memory estimate. The same quantized model can fit at a short context and fail at a long one because KV cache and runtime overhead grow separately.

Read the evidence →
[ Stack Builder — the killer feature ]

Tell us your card.
Get a working stack.

Three inputs — hardware budget, goal, platform — and we assemble a complete, consistent local AI stack from the catalogue. Empty slots tell you why nothing fits, instead of pretending.

Open the console →
Sample output · 16 GB · image gen
  • ENGINEComfyUI8 GB
  • MODELFlux.1 [dev]12 GB
  • RUNNEROllamaCPU+
  • EXTRASComfyUI Managervia host
  • VIDEONo fit at 16 GB — needs 24 GB+

↑ Empty slots are an honest output, not a bug.

[ New — Frontier models ]

Cloud models,
same honest treatment.

25 frontier models with verified per-token pricing, context windows, and sourced benchmarks. Compare any two head-to-head — and when there's a local equivalent, we point you at it.

Browse frontier models →Compare two →
Top 5 by overall benchmark · live pricing

↑ Blended $/1M tokens (3:1 in:out) · benchmarks sourced, never invented.

01 · Start from the goal

What are you trying to make?

Six hand-curated playbooks — each one names the models, the workflow engine, and the VRAM tier you actually need.

All use cases →
PB-01≥ 8 GB

Best tools for cinematic video generation

Local video generation is unusually memory-sensitive. The useful question is not only which card you own, but which model variant, resolution, frame count, precision, and offload strategy you can accept.

Use case · Cinematic video
PB-02≥ 4 GB

Best tools for anime and illustration

Anime and illustration workflows live almost entirely in the SDXL / SD1.5 ecosystem. The model itself matters less than the LoRA + ControlNet stack you build around it.

Use case · Anime & illustration
PB-03≥ 6 GB

Best tools for photoreal portraits and characters

Photoreal portrait work has shifted hard toward Flux since 2024. Whether Flux beats SDXL for you depends on whether you have the VRAM, and whether you need consistent characters across frames.

Use case · Photoreal portraits
PB-04ANY GPU

Best tools for a local AI coding assistant

Running your coding assistant locally keeps your codebase off third-party servers and removes rate-limit anxiety. The right pick is more about the runner than the model — they're all interchangeable at the API layer.

Use case · Local coding assistant
PB-05≥ 12 GB

Best tools for LoRA training and fine-tuning

Training a LoRA is the most demanding thing you can do on a consumer GPU. The trainer matters less than the workflow around it — dataset curation, captioning, and inference testing all live in the same pipeline.

Use case · LoRA & fine-tuning
PB-06≥ 4 GB

Best tools for upscaling, restoration and post-processing

Upscaling and restoration is the most underrated lane in local AI — it's where you can produce genuinely impressive output on entry-level hardware, because the per-frame VRAM cost is low.

Use case · Upscaling & restoration
[ BROWSE THE CATALOGUE ]

Or jump straight to a category

ALL CATEGORIES →
3D & SpatialGenerate 3DGenerate meshes, NeRFs, Gaussian splats, and avatars from images or text.Agent FrameworksBuild AI agentsBuild autonomous AI agents that reason, plan, and use tools.AI Video EditingEdit & upscale videoUpscaling, inpainting, restyle, frame interpolation.Audio & Speech GenerationGenerate audioText-to-speech, voice cloning, music & sound effect generation.Custom Nodes & ExtensionsExtend ComfyUIPlugins, custom nodes, and extensions for ComfyUI & friends.Heavy Image GenerationGenerate imagesHigh-end diffusion models — SDXL, Flux, SD3, Hunyuan.Heavy Video GenerationGenerate videoLocal & cloud video diffusion: Wan, Hunyuan, CogVideoX, LTX.Local LLM RunnersRun LLMs locallyRun large language models on your own hardware.Orchestration & APIsScale workloadsRun pipelines at scale — serverless GPUs, queue managers.Training & Fine-tuningTrain modelsLoRA / DreamBooth / full fine-tune toolkits.Workflow EnginesBuild node graphsNodal & graph-based AI workflow builders.
02 · This week

Trending AI tools

What tinkerers and creators are actually using this week.

New & trending →

ComfyUI

The nodal workflow engine for serious diffusion.

OPEN SOURCE6–16 GB VRAM
VRAM fit6–16 GB

Wan 2.2

Open-weight video diffusion from Alibaba.

OPEN SOURCE12–48 GB VRAM
VRAM fit12–48 GB

Cline

Agentic coding in VS Code — reads, writes, runs, browses.

OPEN SOURCECLOUD · NO GPU

FLUX.1 [dev]

12B parameter open-weight diffusion model.

OPEN SOURCE8–24 GB VRAM
VRAM fit8–24 GB

HunyuanVideo

13B open-weight cinematic text-to-video.

OPEN SOURCE24–48 GB VRAM
VRAM fit24–48 GB
03 · Free & open-weight

Best free / open-weight tools

OSS workflow engines, local runners, and open-weight models. You provide the GPU.

See all →
Actually FreeOpen SourceNo SignupWatermark-Free

ComfyUI

The nodal workflow engine for serious diffusion.

OPEN SOURCE6–16 GB VRAM
VRAM fit6–16 GB

llama.cpp

The C++ inference engine powering most local LLMs.

OPEN SOURCECPU-CAPABLE

FLUX.1 [dev]

12B parameter open-weight diffusion model.

OPEN SOURCE8–24 GB VRAM
VRAM fit8–24 GB

Transformers

The library every LLM ships against first.

OPEN SOURCE2–24 GB VRAM
VRAM fit2–24 GB
04 · Decision tree

Best for…

One tool. One specific situation. No more wading through 50 lookalike entries to find the one that fits your card.

05 · Fresh drops

New AI releases

Freshly launched or recently updated tools.

See all →

Wan 2.2

Open-weight video diffusion from Alibaba.

OPEN SOURCE12–48 GB VRAM
VRAM fit12–48 GB

Kling 2.1

Cloud video gen with strong motion control.

PAID · $10/MOCLOUD · NO GPU

Magi-1

Autoregressive video diffusion at 24 GB.

OPEN SOURCE24–48 GB VRAM
VRAM fit24–48 GB

Hunyuan3D-2

Tencent's open 3D generator — multi-view, PBR, ready-to-use meshes.

OPEN SOURCE12–16 GB VRAM
VRAM fit12–16 GB

NVIDIA Cosmos

World-foundation models for physical AI.

OPEN SOURCE24–80 GB VRAM
VRAM fit24–80 GB
06 · Head-to-head

Side-by-side comparisons

Skip the directory rabbit hole — see the top picks compared directly.

See all →
Updated 2026-05-19VS-01

ComfyUI vs AUTOMATIC1111 vs InvokeAI

Which diffusion workflow engine should you commit to?

VERDICT // ComfyUI wins on flexibility and new-model support. A1111 wins on extension breadth. InvokeAI wins on production polish and canvas-first work.
Updated 2026-05-18VS-02

Wan 2.2 vs HunyuanVideo vs LTX-Video

Local video diffusion at three VRAM tiers.

VERDICT // Wan 2.2 is the new open-weight quality leader. Hunyuan still has the most cinematic motion. LTX-Video is the only one that's practical on a 16 GB card.
Updated 2026-05-15VS-03

Ollama vs LM Studio vs llama.cpp

The three serious ways to run LLMs locally.

VERDICT // Ollama is the fastest path to a working OpenAI-compatible server. LM Studio is the best GUI, especially on Apple Silicon. llama.cpp gives you full control and weirdest quants.
Updated 2026-05-13VS-04

Local video diffusion vs Runway Gen-4

When does paying per credit beat owning a 4090?

VERDICT // For occasional polished shots, Runway is cheaper than buying a workstation GPU. For iteration-heavy work or non-public projects, local Wan/Hunyuan pays for itself in months.
The thesis

Decisions, not directories.

We focus on the few tools worth your time, tell you which are actually free, and pick a working stack for your card — never invent benchmarks. Bookmark us; we update on a rolling basis.

Build your stackHow we pick →