---
title: "Uncensored GGUF LLM models: VRAM per quant vs API price"
canonical: https://uplinkswitch.ai/blog/uncensored-gguf-vs-api
generated: 2026-10-20T21:10:48.911Z
source: https://uplinkswitch.ai/blog/uncensored-gguf-vs-api
base_url: https://api.uplinkswitch.com/v1
---

# Uncensored GGUF models: the VRAM each quant needs, and the same builds per token

Eight uncensored GGUF builds with the file size and VRAM each quant needs, read from the Hugging Face API on 2026-10-19, next to what the same builds cost per token on one key.

October 19, 2026·UplinkSwitch·[All posts](https://uplinkswitch.ai/blog)

Every uncensored GGUF LLM model page ranks the downloads and stops there. This one puts the file size of eight builds, read from the Hugging Face API on 2026-10-19, next to the VRAM each quant needs and what the same builds cost per token on one key. Some you should download. Some you cannot.

## Uncensored GGUF LLM model or API build: who picks which

Download the GGUF if you own a 24 GB card and the model you want is 27B or 31B dense: Qwen 3.8 27B Uncensored and Gemma 4 31B Heretic both fit at Q4\_K\_M, the prompts never leave the machine, and the per-token price is zero. Call the build if it is one of the 3 MoE giants in the table, since 86.7 GB at the smallest quant fits no consumer card, or if you have no GPU at all.

Both tables are below. The first is frozen: file sizes read from the Hugging Face API on 2026-10-19. The second is live: the 11 uncensored, abliterated and roleplay builds on our [uncensored LLM API](https://uplinkswitch.ai/uncensored), priced from the catalog at render, with the day's [real-model check](https://uplinkswitch.ai/proof) verdict per row.

## Uncensored GGUF models: file size and VRAM per quant, 2026-10-19

A GGUF file is one binary with the tensors and the metadata a runner needs ([Hugging Face GGUF docs](https://huggingface.co/docs/hub/gguf), read 2026-10-19). The quant in the file name sets the size: Q4\_K\_M is the common 4-bit pick, Q2 and IQ1 are the 1-to-2-bit files that exist so a 300B model has any chance of loading at all. The table gives each build's Q4\_K\_M file, or for the MoE giants the smallest quant the repo offered, as the Hugging Face API reported it on 2026-10-19.

| Build | Quant | File | Needs about | Card class |
| --- | --- | --- | --- | --- |
| Qwen 3.8 27B Uncensored[JonathanColetti/Qwen3.8-27B-Uncensored-GGUF](https://huggingface.co/JonathanColetti/Qwen3.8-27B-Uncensored-GGUF) | Q4\_K\_M | 16.8 GB | 18.8 GB | 24 GB card |
| Gemma 4 31B Heretic[llmfan46/gemma-4-31B-it-uncensored-heretic-GGUF](https://huggingface.co/llmfan46/gemma-4-31B-it-uncensored-heretic-GGUF) | Q4\_K\_M | 18.7 GB | 20.7 GB | 24 GB card |
| Gemma 4 E4B Uncensored[HauhauCS/Gemma-4-E4B-Uncensored-HauhauCS-Aggressive](https://huggingface.co/HauhauCS/Gemma-4-E4B-Uncensored-HauhauCS-Aggressive) | Q4\_K\_M | 5.3 GB | 7.3 GB | 8 GB card |
| Gemma 4 12B QAT Uncensored[HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced](https://huggingface.co/HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced) | Q4\_K\_M | 7.4 GB | 9.4 GB | 12 GB card |
| Llama 3.3 70B abliterated[bartowski/Llama-3.3-70B-Instruct-abliterated-GGUF](https://huggingface.co/bartowski/Llama-3.3-70B-Instruct-abliterated-GGUF) | Q3\_K\_M | 34.3 GB | 36.3 GB | 48 GB, or two 24 GB cards |
| GLM 5.3 Flash abliterated[huihui-ai/Huihui-GLM-5.3-Flash-abliterated-GGUF](https://huggingface.co/huihui-ai/Huihui-GLM-5.3-Flash-abliterated-GGUF) | UD-IQ1\_S | 93.1 GB | 95.1 GB | 96 GB, workstation or server |
| DeepSeek V4 Flash 0731 abliterated[huihui-ai/Huihui-DeepSeek-V4-Flash-0731-abliterated-GGUF](https://huggingface.co/huihui-ai/Huihui-DeepSeek-V4-Flash-0731-abliterated-GGUF) | Q2 | 86.7 GB | 88.7 GB | 96 GB, workstation or server |
| DeepSeek V4.1 Flash abliterated[audreyt/DeepSeek-V4.1-Flash-Abliterated-GGUF](https://huggingface.co/audreyt/DeepSeek-V4.1-Flash-Abliterated-GGUF) | Q2 | 365.7 GB | 367.7 GB | No consumer card |

File sizes from the Hugging Face API on 2026-10-19, in GB as reported. "Needs about" is a rule of thumb: the file plus 2 GB of KV cache and scratch for an 8K context, all on the GPU. A longer context, a vision projector (the mmproj file) or a draft model adds to it; offloading layers to system RAM subtracts from it at lower speed. Llama 3.3 70B abliterated: The largest quant in our pull; the repo also lists IQ2\_M at 24.1 GB. GLM 5.3 Flash abliterated: The smallest of 47 files; UD-IQ2\_XXS is 101.8 GB. DeepSeek V4 Flash 0731 abliterated: The smallest full-model file; Q4\_K is 164.6 GB. DeepSeek V4.1 Flash abliterated: The only quant the repo offers.

The split is clean. 4 rows fit a single consumer card: Gemma 4 E4B Uncensored at 5.3 GB on an 8 GB card, Gemma 4 12B QAT Uncensored on 12 GB, and the two 27B-to-31B dense builds on 24 GB. Llama 3.3 70B abliterated needs two 24 GB cards at Q3\_K\_M, or an IQ2\_M file at 24.1 GB that leaves no room for context on one. Then the floor jumps: the three MoE builds start at 86.7 GB and DeepSeek V4.1 Flash abliterated is 365.7 GB at its only quant. No card you can buy at retail holds Q2 of it.

Loading one takes one line with [llama.cpp](https://github.com/ggml-org/llama.cpp) (read 2026-10-19): `llama cli -hf <repo>` downloads the GGUF and starts a chat, and `llama serve -hf <repo>` starts an OpenAI-compatible server on it. Ollama and LM Studio load the same files.

## The same uncensored builds per token

The 11 builds below run on one OpenAI-compatible key with no GPU on your side. Where a row matches a GGUF above by name it says so; the rest are builds with no download in the table. Prices are USD per million tokens from the live catalog on 2026-10-20.

| Build | Kind | Input | Output | 100 messages | Daily check |
| --- | --- | --- | --- | --- | --- |
| [GLM-5.3 Flash Uncensored](https://uplinkswitch.ai/models/glm-5.3-flash-uncensored)glm-5.3-flash-uncensoredGGUF above: UD-IQ1\_S, 93.1 GB | Uncensored | $0.084 | $0.084 | $0.0151 | [Same fingerprint since Oct 12](https://uplinkswitch.ai/proof#glm-5.3-flash-uncensored) |
| [GLM-5.3 Uncensored](https://uplinkswitch.ai/models/glm-5.3-uncensored)glm-5.3-uncensored | Uncensored | $0.1746 | $0.1746 | $0.0314 | [Same fingerprint since Oct 12](https://uplinkswitch.ai/proof#glm-5.3-uncensored) |
| [GLM-4.7 Flash Uncensored](https://uplinkswitch.ai/models/glm-4.7-flash-uncensored)glm-4.7-flash-uncensored | Uncensored | $0.12 | $0.60 | $0.036 | [No answer at the last check](https://uplinkswitch.ai/proof#glm-4.7-flash-uncensored) |
| [Roleplay Uncensored](https://uplinkswitch.ai/models/roleplay-uncensored)roleplay-uncensored | Roleplay | $0.30 | $0.60 | $0.063 | [No answer at the last check](https://uplinkswitch.ai/proof#roleplay-uncensored) |
| [GLM-5.3 Flash Abliterated](https://uplinkswitch.ai/models/glm-5.3-flash-abliterated)glm-5.3-flash-abliteratedGGUF above: UD-IQ1\_S, 93.1 GB | Abliterated | $0.66 | $0.66 | $0.1188 | [Same fingerprint since Oct 12](https://uplinkswitch.ai/proof#glm-5.3-flash-abliterated) |
| [DeepSeek V4 Pro Uncensored](https://uplinkswitch.ai/models/deepseek-v4-pro-uncensored)deepseek-v4-pro-uncensored | Uncensored | $0.75 | $0.75 | $0.135 | [Same fingerprint since Oct 12](https://uplinkswitch.ai/proof#deepseek-v4-pro-uncensored) |
| [DeepSeek V4.1 Flash Abliterated](https://uplinkswitch.ai/models/deepseek-v4.1-flash-abliterated)deepseek-v4.1-flash-abliteratedGGUF above: Q2, 365.7 GB | Abliterated | $1.05 | $1.05 | $0.189 | [Same fingerprint since Oct 19](https://uplinkswitch.ai/proof#deepseek-v4.1-flash-abliterated) |
| [Chat Uncensored](https://uplinkswitch.ai/models/chat-uncensored)chat-uncensored | Uncensored | $1.50 | $3.00 | $0.315 | [No answer at the last check](https://uplinkswitch.ai/proof#chat-uncensored) |
| [QwQ 32B Uncensored](https://uplinkswitch.ai/models/qwq-32b-uncensored)qwq-32b-uncensored | Uncensored | $1.98 | $3.00 | $0.387 | [No answer at the last check](https://uplinkswitch.ai/proof#qwq-32b-uncensored) |
| [GLM-5.3 Abliterated](https://uplinkswitch.ai/models/glm-5.3-abliterated)glm-5.3-abliterated | Abliterated | $2.25 | $2.25 | $0.405 | [Same fingerprint since Oct 12, no answer since Oct 20](https://uplinkswitch.ai/proof#glm-5.3-abliterated) |
| [Gemma 31B Heretic Uncensored](https://uplinkswitch.ai/models/gemma-31b-heretic-uncensored)gemma-31b-heretic-uncensoredGGUF above: Q4\_K\_M, 18.7 GB | Uncensored | $1.50 | $8.25 | $0.4725 | [No answer at the last check](https://uplinkswitch.ai/proof#gemma-31b-heretic-uncensored) |

USD per million tokens from the live catalog on 2026-10-20, lowest price per message first. 100 messages means 1,500 tokens in and 300 out each. Daily check verdicts from /proof.

Once a UTC day, every model on the key gets the same fixed test prompt. The input token count works as a tokenizer fingerprint, so a swapped model or hidden instructions added to the prompt shows up as a changed count. At the last check (Oct 20 UTC), 6 of the 11 builds it reached were stable. The [method page](https://uplinkswitch.ai/proof/method) shows the prompt.

Qwen 3.8 27B Uncensored, Gemma 4 E4B Uncensored and Gemma 4 12B QAT Uncensored are not in the catalog today, so for those three the download is the only path. [Gemma 31B Heretic Uncensored](https://uplinkswitch.ai/models/gemma-31b-heretic-uncensored) is the API side of the Gemma 4 31B Heretic row, and [DeepSeek V4.1 Flash Abliterated](https://uplinkswitch.ai/models/deepseek-v4.1-flash-abliterated) is the API side of the Q2 365.7 GB file.

## 100 roleplay messages on both

One roleplay message here is 1,500 tokens in, the character card and the recent chat, and 300 out. On a downloaded GGUF, 100 of them cost nothing per token: the bill is the card, the power it draws while it runs, and the time it takes at your card's speed. On the key, the same 100 messages are priced from the live catalog.

| Path | Build | 100 messages |
| --- | --- | --- |
| Download | Any GGUF above that fits your card | $0.00 per token |
| Call | [GLM-5.3 Flash Uncensored](https://uplinkswitch.ai/models/glm-5.3-flash-uncensored) | $0.0151 |
| Call | [GLM-5.3 Uncensored](https://uplinkswitch.ai/models/glm-5.3-uncensored) | $0.0314 |
| Call | [GLM-4.7 Flash Uncensored](https://uplinkswitch.ai/models/glm-4.7-flash-uncensored) | $0.036 |
| Call | [Roleplay Uncensored](https://uplinkswitch.ai/models/roleplay-uncensored) | $0.063 |
| Call | [GLM-5.3 Flash Abliterated](https://uplinkswitch.ai/models/glm-5.3-flash-abliterated) | $0.1188 |
| Call | [DeepSeek V4 Pro Uncensored](https://uplinkswitch.ai/models/deepseek-v4-pro-uncensored) | $0.135 |
| Call | [DeepSeek V4.1 Flash Abliterated](https://uplinkswitch.ai/models/deepseek-v4.1-flash-abliterated) | $0.189 |
| Call | [Chat Uncensored](https://uplinkswitch.ai/models/chat-uncensored) | $0.315 |
| Call | [QwQ 32B Uncensored](https://uplinkswitch.ai/models/qwq-32b-uncensored) | $0.387 |
| Call | [GLM-5.3 Abliterated](https://uplinkswitch.ai/models/glm-5.3-abliterated) | $0.405 |
| Call | [Gemma 31B Heretic Uncensored](https://uplinkswitch.ai/models/gemma-31b-heretic-uncensored) | $0.4725 |

USD for 100 messages of 1,500 tokens in and 300 out, from the live catalog on 2026-10-20. The download row leaves out hardware and power, which depend on your card.

The spread on the key runs from $0.0151 (GLM-5.3 Flash Uncensored) to $0.4725 (Gemma 31B Heretic Uncensored) per 100 messages. GLM-5.3 Flash Uncensored, whose GGUF row starts at 93.1 GB, costs $0.0151; Gemma 31B Heretic Uncensored, whose GGUF row is 18.7 GB and runs on a 24 GB card, costs $0.4725. A $10 top-up is 66,100 messages on GLM-5.3 Flash Uncensored at this size.

## Where the download wins

-   **Privacy.** The prompt, the character card and the reply stay on your machine. Nothing is sent, logged or billed anywhere.
-   **Zero marginal cost.** Once the card is paid for, the 100th message costs the same as the first: power. A card that already sits in your desktop makes the per-token column above a number you never pay.
-   **A 24 GB card runs a 27B.** Qwen 3.8 27B Uncensored at Q4\_K\_M is 16.8 GB and Gemma 4 31B Heretic is 18.7 GB on 2026-10-19. Both leave room for an 8K context on 24 GB by the rule here.
-   **The exact file.** You pick the repo, the quant and the sampler, and the model does not change under you. Qwen 3.8 27B Uncensored, Gemma 4 E4B Uncensored and Gemma 4 12B QAT Uncensored have no API twin here today, so the download is the only way to run them.
-   **Offline.** A GGUF on a laptop works on a plane.

## Where the API wins

-   **The 300B-class MoE builds.** GLM 5.3 Flash abliterated (93.1 GB at UD-IQ1\_S), DeepSeek V4 Flash 0731 abliterated (86.7 GB at Q2) and DeepSeek V4.1 Flash abliterated (365.7 GB at Q2) fit no consumer card at any quant in their repos. On the key, DeepSeek V4.1 Flash Abliterated and GLM-5.3 Flash Uncensored run at 300\-token replies with nothing to download.
-   **No GPU.** A laptop with integrated graphics, a server with none, a Raspberry Pi: the model runs behind the API and your code sends HTTP.
-   **Phones.** A roleplay frontend on a phone cannot hold a 17 GB file in memory. It can hold a base URL and a key.
-   **Switching builds.** 11 builds are one id apart. Trying a second one is a string change, not a 19 GB download.
-   **The daily check.** Each row's verdict is in the table and on [/proof](https://uplinkswitch.ai/proof), updated once a UTC day, so you can see the same model answered the same way as yesterday.
-   **One key for the rest.** The same key runs Claude, GPT and Gemini. A local runner runs what is on disk.

## How to switch from Ollama or llama.cpp

If you already run an Ollama uncensored model through its OpenAI-compatible endpoint, the switch is the base URL, the key and the model id. Ollama's endpoint is `http://localhost:11434/v1` and it ignores the key (Ollama's docs, read 2026-10-19); llama.cpp's `llama serve` is the same shape. Ours is `https://api.uplinkswitch.com/v1`, and the key is real.

python

```
from openai import OpenAI

client = OpenAI(
    base_url="https://api.uplinkswitch.com/v1",  # was http://localhost:11434/v1
    api_key="sk-uplink-live-...",  # was "ollama"
)

completion = client.chat.completions.create(
    model="deepseek-v4.1-flash-abliterated",  # was a local GGUF tag
    max_tokens=600,
    messages=[
        {"role": "system", "content": "You are Vex, the villain of a dark fantasy novel."},
        {"role": "user", "content": "Threaten the captured knight. Stay in character."},
    ],
)
print(completion.choices[0].message.content)
```

The SDK sends the key in an `Authorization: Bearer` header; `x-api-key` works too, and a key in the URL is refused. Streams pass through unchanged, so streaming code needs no edit. One request lists every id the key can call:

shell

```
curl https://api.uplinkswitch.com/v1/models -H "Authorization: Bearer $UPLINKSWITCH_API_KEY"
```

SillyTavern takes the base URL ending in `/v1` as a Custom (OpenAI-compatible) source, the same way it takes a local runner. The [SillyTavern guide](https://uplinkswitch.ai/guides/sillytavern) has every field. GLM-5.3 Uncensored was the most-used build here by paid spend on 2026-10-18, and the abliterated builds think before they answer, so give them 1,000 or more max tokens. What abliteration changes in the weights, GGUF or served, is in [what an abliterated model is](https://uplinkswitch.ai/blog/abliterated-models).

## Questions

**Which uncensored GGUF model fits a 6 GB card?** By the rule of thumb here (file plus about 2 GB for an 8K context), none of the Q4\_K\_M files in the table. The closest is Gemma 4 E4B Uncensored from HauhauCS/Gemma-4-E4B-Uncensored-HauhauCS-Aggressive: its Q4\_K\_M was 5.3 GB on 2026-10-19, and the same repo listed an IQ3\_M at 4.7 GB and a Q2\_K\_P at 4.4 GB. On a 6 GB card, run one of those and let the runner put the rest of the layers in system RAM, at lower speed.

**What uncensored GGUF runs in 8 GB of VRAM?** Gemma 4 E4B Uncensored at Q4\_K\_M (5.3 GB on 2026-10-19) is the one row in the table that lands in the 8 GB class with room for context. Gemma 4 12B QAT Uncensored at Q4\_K\_M is 7.4 GB, which the rule puts at a 12 GB card; on 8 GB it runs with some layers offloaded.

**What is the best uncensored LLM GGUF for a 24 GB card?** Qwen 3.8 27B Uncensored at Q4\_K\_M (16.8 GB) and Gemma 4 31B Heretic at Q4\_K\_M (18.7 GB) both fit a 24 GB card with an 8K context by the rule here. We have not benchmarked either, so pick by the base model you want. Llama 3.3 70B abliterated needs two 24 GB cards at Q3\_K\_M, or a 2-bit file on one.

**What are the best GGUF models for roleplay?** For a download, the consumer-card rows above: Qwen 3.8 27B Uncensored, Gemma 4 31B Heretic, Gemma 4 E4B Uncensored and Gemma 4 12B QAT Uncensored. For a call, GLM-5.3 Flash Uncensored is the lowest price per message here at $0.0151 per 100 messages, and the roleplay and DeepSeek builds were run in SillyTavern on 2026-10-18. See [Best AI for roleplay](https://uplinkswitch.ai/blog/best-ai-for-roleplay).

**Can Ollama run an uncensored model?** Yes. Ollama loads GGUF files, and its OpenAI-compatible endpoint is http://localhost:11434/v1 (Ollama's docs, read 2026-10-19). A 27B Q4\_K\_M file such as Qwen 3.8 27B Uncensored needs a 24 GB card to stay on the GPU. The 3 MoE builds in the table are 86.7 GB and up at their smallest quant, so they do not run on a consumer card under Ollama or anything else.

**Is there an uncensored LLM leaderboard?** Not a quality one here; we have not run a benchmark. The live table on this page orders 11 builds by price per message, each with the day's real-model verdict from /proof. See [Best uncensored LLM](https://uplinkswitch.ai/blog/best-uncensored-llm).

**Where do GGUF uncensored models come from?** From Hugging Face repos, most of them community conversions: the eight rows above are from JonathanColetti, llmfan46, HauhauCS, bartowski, huihui-ai and audreyt. The Hub tags GGUF files and shows their metadata and tensor shapes in a viewer (Hugging Face GGUF docs, read 2026-10-19).

**Are the GGUF builds and the API builds the same model?** 3 rows match a catalog id by name: Gemma 4 31B Heretic (gemma-31b-heretic-uncensored), GLM 5.3 Flash abliterated (glm-5.3-flash-abliterated, glm-5.3-flash-uncensored) and DeepSeek V4.1 Flash abliterated (deepseek-v4.1-flash-abliterated). The catalog does not publish which repo or quant serves each id, so treat them as the same base model changed the same way, not as byte-identical weights. 4 rows, Gemma 4 E4B Uncensored, Gemma 4 12B QAT Uncensored, Llama 3.3 70B abliterated and DeepSeek V4 Flash 0731 abliterated, are download-only today.

Sources: the [Hugging Face GGUF docs](https://huggingface.co/docs/hub/gguf) and the [llama.cpp README](https://github.com/ggml-org/llama.cpp), both read 2026-10-19; file sizes from the Hugging Face API on 2026-10-19.

[Get a key](https://uplinkswitch.ai/console): prepaid from $10, no subscription, 30-day price lock.

## Related

-   [Uncensored LLM API](https://uplinkswitch.ai/uncensored)
-   [Best uncensored LLM](https://uplinkswitch.ai/blog/best-uncensored-llm)
-   [Abliterated models explained](https://uplinkswitch.ai/blog/abliterated-models)
-   [Uncensored Gemma API](https://uplinkswitch.ai/blog/uncensored-gemma-heretic)
-   [SillyTavern setup](https://uplinkswitch.ai/guides/sillytavern)
-   [GLM-5.3 Uncensored](https://uplinkswitch.ai/models/glm-5.3-uncensored)
-   [DeepSeek V4.1 Flash Abliterated](https://uplinkswitch.ai/models/deepseek-v4.1-flash-abliterated)
-   [Gemma 31B Heretic Uncensored](https://uplinkswitch.ai/models/gemma-31b-heretic-uncensored)
-   [Daily real-model check](https://uplinkswitch.ai/proof)