Three machines, one memory pool
You no longer need a data centre to run a capable AI model. A top-specification MacBook Pro or NVIDIA's small DGX Spark box runs models that, two years ago, needed a server rack. Almost all of them come from one place: Hugging Face, the public library where AI labs and independent builders publish model weights.
This note does two things. It lists the regular models people actually download to run on these machines, ranked by Hugging Face's own download counts. Then it compares them with their abliterated twins: copies with the refusal behaviour taken out.
The figure that decides what you can run is memory. The whole model has to fit in it. All three machines share one large pool of memory between the processor and the graphics chip, which is why they are good at this.
| Machine | Most memory | Memory bandwidth | Where it shines |
|---|---|---|---|
| MacBook Pro, M4 Max | 128 GB | up to 546 GB/s | Fast, quiet, portable chat and writing |
| MacBook Pro, M5 Max | 128 GB | up to 614 GB/s | The same idea, a little quicker |
| NVIDIA DGX Spark | 128 GB | 273 GB/s | NVIDIA's software stack; long documents; two can be linked |
Bandwidth figures are the fastest configuration of each chip, from Apple's M4 Max specifications, Apple's M5 Max specifications and NVIDIA's DGX Spark hardware guide.
In plain terms, all three hold about the same size of model. The Macs usually write words faster, because their memory moves data faster. The Spark is stronger at reading in a large amount of text before it answers, and it runs the NVIDIA tools most AI software is built around. I measured the same split between a Mac and an NVIDIA card in Same model, two machines.
The rule of thumb for what fits
Most people download a compressed 4-bit version of a model rather than the full-size original. In that form a model needs roughly 0.6 GB of memory for every billion parameters, plus room for the conversation itself.
- A 27-billion-parameter model needs about 16 GB.
- A 120-billion-parameter model needs about 65 GB.
- Anything much past 180 billion will not fit on one 128 GB machine.
That is an estimate, not a guarantee. Different formats and long conversations shift it. How I choose a quantization covers the trade-offs properly.
The popular regular models
These are the most downloaded general-purpose models that a 128 GB machine can realistically run, with Hugging Face's rolling 30-day download counts as of 2 October 2026. Counts move daily; the order is the useful part.
| Model | Size | Downloads, 30 days | Fits in 128 GB? | Why people pick it |
|---|---|---|---|---|
| Gemma 4 26B-A4B | 26B, about 4B active at once | 13.1 M | Easily | The most downloaded here; fast and well-rounded |
| Gemma 4 31B | 31B | 9.9 M | Easily | The more thoughtful, slower Gemma |
| Qwen 3.8 27B | 28B | 6.9 M | Easily | Today's all-rounder; strong at code and reasoning |
| gpt-oss-20b | 21B | 6.7 M | Easily | OpenAI's open model; light and dependable |
| Ornith 1.5 9B | 9B | 5.1 M | Easily | Small and quick for everyday tasks |
| gpt-oss-120b | 117B | 4.5 M | Yes, with room to spare | The largest model most people run comfortably on one machine |
| Qwen 3.6 35B-A3B | 36B, about 3B active | 3.2 M, plus 6.3 M for NVIDIA's Spark-ready build | Easily | Very fast; a DGX Spark favourite |
| Qwen 3.8 Flash-Next | 180B | 1.3 M | Only just | Near the ceiling; little room left for long chats |
| DeepSeek V4 Flash | 304B | 4.5 M | Not on one machine | Popular, but needs two linked Sparks or a bigger workstation |
“Active” matters for speed, not for memory. A model built from many smaller experts only uses a few of them for each word, so it answers quickly, but every expert still has to sit in memory.
Which download to pick on each machine
The same model is usually published several times in different file formats. Picking the right one matters more than most people expect.
- On a Mac: look for MLX versions, such as the mlx-community uploads, or GGUF files for apps like LM Studio and Ollama.
- On a DGX Spark: look for NVFP4 versions, such as NVIDIA's own Qwen 3.6 build. They are made for the Spark's chip and run noticeably better on it.
If you want a single starting point, try Qwen 3.8 27B or Gemma 4 31B as the everyday model, and gpt-oss-120b when you want the most capable model that still fits comfortably. I run Qwen 3.8 27B on my own M4 Max; Qwen 3.8 moved the frontier onto my Mac has the details, and Qwen3.8-Flash-Next system requirements covers the 180B model at the edge of what fits.
What “abliterated” means
Browse Hugging Face for long and you will see names like Qwen3.8-27B-abliterated, gemma-4-31B-it-abliterated or Qwen3.8-27B-Heretic-Abliterated sitting next to the originals. Some are downloaded almost as often.
The short version: an abliterated model is a regular model with its “I can't help with that” reflex removed.
How that works, without the maths: researchers found that when a model decides to refuse, that decision mostly travels along one identifiable path inside it. The finding is in the 2024 paper Refusal in Language Models Is Mediated by a Single Direction. Abliteration locates that path and switches it off. Nothing is retrained and no knowledge is added; the model simply stops steering itself towards a refusal. Heretic is a popular open-source tool that does this automatically.
One naming caution. “Uncensored” on a model card can mean abliteration, extra fine-tuning, or both. Read the card before assuming which method was used.
The popular abliterated versions
| Abliterated model | Based on | Downloads, 30 days |
|---|---|---|
| Huihui Qwen3.8-27B abliterated (GGUF) | Qwen 3.8 27B | 2.0 M |
| Qwen3.8-27B Heretic abliterated (GGUF) | Qwen 3.8 27B | 1.4 M |
| OBLITERATUS Qwen3.8-27B | Qwen 3.8 27B | 0.8 M |
| Qwen3.6-35B-A3B abliterated | Qwen 3.6 35B-A3B | 0.6 M |
| Huihui DeepSeek V4 Flash abliterated (GGUF) | DeepSeek V4 Flash | 0.4 M |
| Gemma 4 31B abliterated | Gemma 4 31B | 0.3 M |
| Huihui Qwen3.8 Flash-Next abliterated (GGUF) | Qwen 3.8 Flash-Next | 0.2 M |
| gpt-oss-20b abliterated (several uploads) | gpt-oss-20b | under 0.1 M combined |
Two patterns stand out. Qwen dominates the abliterated world, while OpenAI's gpt-oss models are rarely abliterated. And an abliterated copy is the same size as its original: if the regular model fits your machine, so does its twin, at the same speed.
Four ways the Qwen versions are made
Every abliterated Qwen is the same model with its refusal behaviour removed, but publishers go about it in quite different ways, and that shows up in quality. Here is what the most downloaded Qwen 3.8 versions say about themselves on their own model pages.
1. The original method. Huihui's Qwen3.8-27B and its Flash-Next sibling use the classic recipe: find the refusal pathway and switch it off in one go. The page itself calls the method crude and proof-of-concept, and publishes no test results. It is the most popular abliterated build, but popularity is the only evidence on offer.
2. The automated method. The tool Heretic searches for the setting that removes the most refusals while damaging the model the least. The most downloaded “Uncensored” Qwen 3.8, about 2.2 million downloads in 30 days, is made this way and has the most transparent page I found. Refusals fell from 98 to 12 out of 100 test prompts, while its general-knowledge score barely moved (83.4 to 83.3), and the author states plainly that refusals are reduced, not eliminated. The Heretic-Abliterated version runs the process three times and reports zero or one refusal per 100, yet deliberately leaves a few guardrails in place.
3. The blended method. OBLITERATUS combines several removal techniques and claims zero refusals across more than a thousand prompts. Its own page shows the cost: the same general-knowledge test drops from 84.5 to 82.3. It also warns that adding a system prompt can bring refusals back.
4. Repackages. Some popular files are someone else's uncensored model, resized or rebuilt. The Mac-friendly MLX version is a compressed copy of another publisher's model, and its page does not say how the refusals were removed. Another build splices parts of an uncensored model into a smaller file for 12 GB graphics cards, and says openly that refusals were not measured.
| Method | Example | Refusals left, as published | Quality, as published |
|---|---|---|---|
| Original | Huihui Qwen3.8-27B | Not measured | Not measured |
| Automated (Heretic) | Qwen3.8-27B Uncensored | 12 in 100 | 83.4 → 83.3 |
| Automated, three passes | Qwen3.8-27B Heretic-Abliterated | 0–1 in 100 | Small drift, by its own measure |
| Blended | OBLITERATUS Qwen3.8-27B | 0, claimed | 84.5 → 82.3 |
| Repackage | MLX Uncensored 4-bit | Not stated | Not stated |
The pattern across these pages: the more completely refusals are stripped out, the more quality tends to go with them. The name tells you almost nothing. Before downloading one, look for three things on the model page: a refusal count from a real test set, a benchmark compared with the original model, and an honest list of limitations. Pages that test on ten prompts, or not at all, are asking you to take them on trust.
Regular vs abliterated: the real differences
| Question | Regular model | Abliterated model |
|---|---|---|
| Does it refuse? | Sometimes, including harmless requests it misreads, like “how do I kill a Python process?” | Rarely or never |
| Does it know more? | No | No. Removing refusals adds no knowledge |
| Answer quality | The lab's tested release | Often slightly lower; reasoning and accuracy can slip, more so with rougher methods |
| Safety net | Built in | Gone; judging what is appropriate is entirely on you |
| Who made it | The original lab, such as Google, Alibaba or OpenAI | Usually an individual or small group; quality and trust vary |
| Licence | The lab's licence | Still the lab's licence; abliteration does not change your obligations |
| Speed and memory | Baseline | Identical |
Why people use them, and why you might not
The reasons people choose them:
- Fewer false refusals. Regular models sometimes block perfectly legitimate requests in fiction, security research, medicine or history. This is the most common reason.
- Research and testing. Security teams and researchers use them to see how a model behaves without its guardrails.
- Control. On their own machine, some people simply prefer to set the rules themselves.
The reasons to think twice:
- It is not smarter. “Uncensored” in a name sounds like extra ability. It is the same model with fewer brakes.
- It is less polished. Expect a few more confident wrong answers and a little less care.
- Responsibility moves to you. Nothing stops harmful or inappropriate output, so keep it away from children, customers and anything public-facing.
- Check the uploader. Prefer well-known publishers with many downloads and a clear model card, and the safer safetensors or GGUF file formats.
I go further into what local and uncensored models are good for in Local & uncensored models.
The bottom line
- On a 128 GB M4 Max, M5 Max or DGX Spark, the popular sweet spot is a 27 to 35 billion parameter model such as Qwen 3.8, Gemma 4 or Qwen 3.6 for everyday speed, or gpt-oss-120b for the most capability on one machine.
- Mac owners want MLX or GGUF downloads. Spark owners want NVFP4.
- Abliterated models are the same models with refusals switched off: the same size and knowledge, slightly less polish and no safety net. They are useful for some jobs but they are not an upgrade. Start with the regular version, and switch only if refusals actually get in the way.
How these figures were gathered. Download counts, parameter counts and model names come from the Hugging Face model API on 2 October 2026. Hardware figures come from the Apple and NVIDIA specification pages linked above. The memory rule of thumb is an approximation, not a measurement.