Live trending boards for models, datasets and Spaces — pulled straight from the public Hugging Face API by your own browser — plus deep, honest guides to actually running those models on your own hardware. No cloud lock-in, no fake numbers.
abuz8 account currently has no public models, datasets or Spaces — and we won't pretend otherwise.
Publishing sovereign models and datasets is on our roadmap.
The 24 models trending on Hugging Face right now, sorted by trending score. Each card links out to the model's own page — the authors deserve the visit.
Data: Hugging Face public API (/api/models?sort=trendingScore), fetched client-side. Download and like counts belong to each model's authors. Not affiliated with Hugging Face.
Trending public datasets — the raw material of open AI. Same live feed, same rule: every number comes from the Hugging Face API in real time or the board says so.
Data: Hugging Face public API (/api/datasets?sort=trendingScore), fetched client-side. All datasets belong to their authors and carry their own licenses — check each dataset card before use.
Spaces are live demo apps hosted on Hugging Face — the fastest way to try a model before downloading anything. These are trending right now.
Data: Hugging Face public API (/api/spaces?sort=trendingScore), fetched client-side. Each Space is built and hosted by its author under its own license.
This is the sovereign-AI part: taking an open model off the hub and running it on your hardware — no API key, no cloud bill, no data leaving your machine. Everything below is public knowledge, tested on our own RTX 5090 rig.
Ollama is the fastest path (one installer, one command). llama.cpp gives maximum control. LM Studio is the point-and-click option. Comparison table below.
On huggingface.co/models, filter by the GGUF library tag, or search a model name plus "GGUF". Community quantizers publish ready-to-run files for most popular open models.
Use the quantization and VRAM tables below. Default pick: the Q4_K_M file that fits your GPU with a gigabyte or two to spare.
One command — see the code block underneath these steps. First run downloads or loads the model; after that it's instant and fully offline.
All three runners expose an OpenAI-compatible local endpoint (Ollama: localhost:11434/v1 · llama-server: localhost:8080/v1 · LM Studio: localhost:1234/v1), so any tool that speaks the OpenAI API can point at your own machine instead.
# Ollama — easiest (after installing from ollama.com) ollama run <model-name> # llama.cpp — OpenAI-compatible local server llama-server -m ./your-model-Q4_K_M.gguf -c 4096 --n-gpu-layers 99
GGUF is the single-file model format used by llama.cpp and everything built on top of it. One .gguf file bundles the weights, tokenizer and metadata; it supports quantized weights and is memory-mapped, so models load fast and don't need unpacking. When you see "GGUF" on a Hugging Face repo, it means: ready to run locally with the tools on this page.
Quantization stores weights in fewer bits, shrinking the file and the VRAM needed, at some cost in quality. Sizes below are approximate for an 8-billion-parameter model and vary by architecture.
| Quant | ~Bits / weight | ~File size (8B model) | Quality impact | When to use |
|---|---|---|---|---|
Q2_K | ~2.6 | ~3.2 GB | Severe — often noticeably degraded | Only when nothing else fits |
Q3_K_M | ~3.9 | ~4.0 GB | Noticeable loss | Very tight VRAM budgets |
Q4_K_M | ~4.8 | ~4.9 GB | Small loss — the community default | Best size/quality balance for most people |
Q5_K_M | ~5.7 | ~5.7 GB | Very small loss | When you have VRAM headroom |
Q6_K | ~6.6 | ~6.6 GB | Barely measurable loss | Quality-first on 12 GB+ cards |
Q8_0 | ~8.5 | ~8.5 GB | Practically indistinguishable | Maximum quality, archival |
F16 | 16 | ~16 GB | Original half precision | Fine-tuning bases, reference runs |
| llama.cpp | Ollama | LM Studio | |
|---|---|---|---|
| What it is | The core C/C++ inference engine GGUF runs on | A model runner built around llama.cpp | Desktop GUI for finding + running GGUF models |
| Interface | CLI + llama-server | CLI + background service | Point-and-click app |
| Install | Build from source or grab release binaries | One installer | One installer |
| Model management | Manual — you download .gguf files yourself | ollama pull from its registry | Built-in Hugging Face model browser |
| OpenAI-compatible API | Yes (llama-server, port 8080) | Yes (port 11434, plus its own API) | Yes (local server mode, port 1234) |
| Open source | Yes — MIT | Yes — MIT | App is free but closed-source |
| Best for | Maximum control, servers, tinkering | Fastest path to a running model | Visual browsing, non-technical users |
llama.cpp by Georgi Gerganov & contributors · Ollama by Ollama Inc. · LM Studio by Element Labs — all credited to their authors; none are ABUZ8 products.
Rule of thumb: GGUF file size + 1–2 GB for the context cache and overhead. Longer context windows eat more VRAM. llama.cpp-family runners can split layers between GPU and CPU RAM (--n-gpu-layers) — slower, but it lets a big model run on a small card.
| GPU VRAM | Comfortable fit (rule of thumb) |
|---|---|
| 4 GB | 1–3B models at Q4 |
| 6–8 GB | 7–8B at Q4_K_M |
| 12 GB | 8B at Q8, or 13–14B at Q4 |
| 16 GB | 14B at Q6, or ~20B class quantized |
| 24 GB | ~30B class at Q4–Q5 |
| 32 GB | ~30B at Q6–Q8, or 70B at low quant with partial CPU offload |
A hub full of model repos and one question: which one should you run? Work through license first, size second, benchmarks last — in that order.
"Open weights" is not one thing. The model card and LICENSE file on the model's own Hugging Face page are the source of truth — licenses vary by model and by version, so read the actual file before you build on anything.
| License family | Commercial use | What to watch |
|---|---|---|
| Apache-2.0 / MIT | Yes | Genuinely permissive — use, modify, redistribute |
| Custom "community" licenses (e.g. Llama-style) | Usually, with conditions | Extra terms such as usage policies or scale thresholds — read the license text itself |
| Open-weights with usage terms (e.g. Gemma-style) | Conditional | Acceptable-use policies attached to the weights |
| Non-commercial / research-only | No | Fine for learning and evaluation, not for products |
| Gated models | Varies | Requires a Hugging Face account + accepting the author's terms before download |
Straight answers — including the ones that don't flatter us.
abuz8 account exists but currently has zero public models, datasets or Spaces — and we won't pretend otherwise. Publishing sovereign models and datasets is on our roadmap..gguf file contains the weights, tokenizer and metadata, supports quantized weights, and is memory-mapped for fast loading.Q4_K_M is the widely used default balance of size and quality. Pick Q5_K_M or Q6_K if you have VRAM headroom, and treat Q2/Q3 quants as a last resort — the quality loss becomes noticeable, especially on smaller models. Full table in the Run Locally tab.