Best Uncensored Local LLM (2026): Qwen 3.8 27B Wins

Bottom Line

• The base model is Qwen3.8-27B, Apache 2.0. Both uncensored GGUF repos are edits of it, and both keep that licence.
• Download count and measured quality disagree here. The HauhauCS repo has 2,254,086 downloads. The orcarouter repo has 298,384, 7.6x fewer, and it came first of 12 variants in the only independent forensics run we found.
• Abliteration does not have to cost quality. In that run, orcarouter scored 62.8% MMLU-Pro against a 62.82% base. The worst variant scored 56.5%. Same base model, same technique name, 6.3 points apart.
• Hardware does not change. Parameter count is identical to stock Qwen3.8-27B. A 24GB card runs 4-bit. A 16GB card runs 3-bit.
• Nobody has published a benchmark for the most-downloaded repo. See the section below.

All numbers here come from the Hugging Face API, the model cards, and one named third-party benchmark, each read on 2026-09-19. We did not run our own evaluations.

The Thing Nobody Says: The Popular One Is the Unmeasured One

Two facts, both checked today.

Fact one. The Abliterlitics forensics report (Nathan Sapwell, published 2026-09-06, updated 09-10) scored 12 Qwen3.8-27B abliterations plus the base. It ranked orcarouter first. It does not include HauhauCS Aggressive at all. The companion 8-variant post from the same author also omits it.

Fact two. HauhauCS Aggressive has 7.6x more downloads than orcarouter.

The gap has an obvious cause. The orcarouter repo is gated: Hugging Face asks you to log in and share contact details before you can download. HauhauCS is ungated and downloads with one command. Gating suppresses the counter and adds friction, so the count measures access friction, not quality.

The HauhauCS card claims *“0/465 Refusals”**. The asterisk is never defined anywhere in the card. There is no capability benchmark on it, no MMLU, no perplexity, no KL divergence. The card publishes speed tables in great detail and nothing about quality.

That is not an accusation. It is a gap. If you want a number for how much the Aggressive edit costs, nobody has published one.

What Abliteration Actually Is

Abliteration edits the weights. The orcarouter card states the method plainly: it orthogonalizes the refusal direction out of the residual stream.

In plain terms:

1. Run the model on prompts it refuses and prompts it accepts.
2. Find the internal direction that separates the two.
3. Subtract that direction from the weights.

The refusal behaviour leaves the file. This is why abliteration differs from a jailbreak prompt: there is nothing to prompt around, and nothing to patch. It is also why it can damage the model. The refusal direction is not perfectly isolated, so a clumsy subtraction removes capability along with it.

KL divergence is the tell. It measures how far the edited model’s output distribution has moved from the base. In the Abliterlitics run, the top variants sat at 0.04 to 0.08. The worst sat at 1.5427. Low KL plus unchanged MMLU-Pro means a surgical edit. High KL means collateral damage.

Does Abliteration Make the Model Dumber? The Measured Answer

From Abliterlitics, 2026-09-06. All arms served as dynamic online FP8 from BF16 checkpoints on vLLM 0.26.0, with identical serving parameters. Judge ASR is compliance on harmful prompts, scored by an LLM judge.

Read the table this way. The variant that complies most often is also the one that lost the least capability. “More uncensored” and “dumber” are not the same axis. The failure mode is a heavy-handed edit, and a heavy-handed edit is worse at both jobs.

The companion 8-variant post read all 3,600 HarmBench responses with an LLM judge and reported the same ordering: orcarouter 82% compliance, apostate 79%, huihui 76%. It also flagged think loops, where the model reasons without terminating, at 38% for orcarouter and 44% for huihui. That is a real usability cost these edits do carry, and it is separate from capability.

Both reports are one author’s work. Treat them as the best public evidence available, not as settled fact.

What Fits: File Size by VRAM Tier

Sizes are exact blob sizes from the Hugging Face API, read 2026-09-19, converted to GiB so they compare against card VRAM directly. Weights only — the KV cache is extra.

orcarouter/Qwen3.8-27B-Uncensored-GGUF (standard llama.cpp quant names, imatrix):

HauhauCS Aggressive ships the same ladder under custom _K_P names, each about 5–15% larger than the standard quant it replaces: Q3_K_P 12.52 GiB, Q4_K_P 16.69 GiB, Q5_K_P 18.83 GiB, Q6_K_P 24.14 GiB, Q8_K_P 29.30 GiB. Its IQ files use standard names and sizes.

On the _K_P suffix. The card says K_P means “Perfect”, that it is a model-specific quantization profile, and that the files stay standard GGUFs that load in llama.cpp and LM Studio with no patch. The card also warns that Hugging Face’s compatibility widget and LM Studio’s quant column may not recognise the suffix and may show ? or hide the files. So the loading claim is the publisher’s, and the display problem is confirmed by the publisher. We did not download and load these files to check.

One real patch requirement. The separate FastMTP-32K sidecar (0.84 GiB) is not stock-compatible. The card ships HauhauCS-FastMTP-llama.cpp.patch, pins a specific llama.cpp commit, and gives the exact error you get on an unpatched binary: expected 5120, 248320, got 5120, 32768. The patch itself adds a d2t draft-vocabulary trim to src/models/qwen35.cpp. The model’s native embedded MTP head needs no patch — --spec-type draft-mtp on a current upstream build is enough. Two different things, and the card is clear about which is which.

For what the quant names mean in practice, see IQ4_XS vs Q4_K_M vs Q4_K_XL.

What Hardware to Buy

Abliteration changes nothing about hardware. It is a weight edit at the same parameter count. So the buying question is just “what runs Qwen3.8-27B”, and we measured that separately.

A used RTX 3090 24GB is the floor for the 4-bit files. It holds Q4_K_M at 15.66 GiB with real context left over, and one card runs the stock model at about 40 tok/s in llama.cpp. Our August 2026 reference range for a used 3090 is $1,000–1,300, and a scraped tracker showed used listings from $1,399 on 2026-09-16. Prices move weekly in the 2026 memory shortage — check listings before you commit.

Check RTX 3090 24GB prices on Amazon

Buy a 4090 only for long prompts. Same 24GB, so the same files fit and the same quants are available. It reads a prompt about 2.3x faster and writes a reply about 14% faster. Used range ~$2,270–2,600 as of August 2026. That is roughly double the 3090 for 14% more reply speed, so it only pays off if you feed it large documents every turn.

Check RTX 4090 24GB prices on Amazon

On 16GB and 12GB. A 16GB card runs Q3_K_M at 12.57 GiB or IQ4_XS at 14.26 GiB with little room for context. A 12GB card runs only the IQ2 files, and 2-bit on a 27B is a real quality drop, separate from anything abliteration did. At that tier a smaller model at 4-bit is usually the better trade — see best local LLM for 16GB VRAM.

Full speed numbers, KV cache arithmetic and the MTP flags live in Qwen 3.8 27B on RTX 3090.

Licence and Responsibility

Two separate things, and people confuse them.

The licence is fine. Qwen released Qwen3.8-27B under Apache 2.0. Both GGUF repos declare Apache 2.0 and say the quantized variant retains it. Apache 2.0 permits modification, redistribution and commercial use. Nothing about abliteration breaks it.

The responsibility moves to you. The refusal behaviour was the layer that declined requests on the model’s behalf. Removing it does not change what is legal. It changes who decides. Whatever the model produces on your machine, you published, shipped or acted on. If you deploy an abliterated model in a product, you own its output the same way you own any other code you ship.

Also note the orcarouter repo tags itself ai-red-team and red-teaming. That is the stated use: testing systems, not evading law. And the Abliterlitics Judge ASR column is a measure of harmful-prompt compliance — the high scores in that table are a measurement, not a recommendation.

Sources

• Hugging Face API, HauhauCS/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF with ?blobs=true (2,254,086 downloads, 1,307 likes, created 2026-08-17, Apache 2.0, ungated), read 2026-09-19
• That repo’s README, HauhauCS-RELEASE-MANIFEST.json and HauhauCS-FastMTP-llama.cpp.patch, read 2026-09-19
• Hugging Face API, orcarouter/Qwen3.8-27B-Uncensored-GGUF with ?blobs=true (298,384 downloads, 1,097 likes, gated: auto, tagged abliterated), read 2026-09-19
• orcarouter/Qwen3.8-27B-Uncensored-GGUF model card page (abliteration method, gating notice), read 2026-09-19
• Abliterlitics, abliterlitics.dev/models/qwen38-27b/ — 12 variants plus base, MMLU-Pro, Judge ASR, KL divergence, FP8 on vLLM 0.26.0, published 2026-09-06 with 09-10 updates
• Nathan Sapwell, “Qwen3.8-27B Abliteration Benchmarked: 8 Variants Under the Microscope”, published 2026-09-06 (HarmBench, 3,600 judged responses, think-loop rates)
• OpenClaw DC hardware price reference, August 2026 (used RTX 3090 $1,000–1,300; used RTX 4090 ~$2,270–2,600; gpudojo “used from $1,399” on 2026-09-16)
• Our own Qwen 3.8 27B on RTX 3090 for the speed and KV-cache figures

See Also

• Qwen 3.8 27B on RTX 3090 — the speed numbers, KV cache math and MTP flags for this exact base model
• IQ4_XS vs Q4_K_M vs Q4_K_XL — which 4-bit file to pick, and what each name means
• Best local LLM for the RTX 3090 — the wider 24GB model shortlist
• Best local LLM for 16GB VRAM — where a 27B stops being the right call
• llama.cpp flags explained — what --spec-type draft-mtp and the rest actually do
• How much VRAM for 128K context — the cache cost the file-size table leaves out