DefiledAI Community

MODEL REVIEWS

Community reviews of uncensored and abliterated models. Structured format: hardware, use case, verdict, pros and cons.

5
Reviews
4.6 / 5
Avg Rating
5
Models Covered
Llama 3.1 70B AbliteratedABLITERATED
2× RTX 3090 NVLink · Q4_K_M · Creative writing / roleplay
neuralrig · 2026-05-28

The best locally-runnable model for unrestricted creative tasks. Post-abliteration quality is indistinguishable from base in creative output. No refusals, no moralizing, just clean output.

Pros
Near-zero quality loss, handles long-form fiction well, instruction following intact
Cons
Needs 48GB VRAM, 21 tok/s is usable but not fast
Dolphin 2.8 Mistral 7BDOLPHIN
RTX 3080 10GB · Q4_K_M · General chat
localai_fan · 2026-05-27

The best entry-level uncensored model. Fast, coherent, and genuinely useful for everyday tasks. The Dolphin fine-tune adds helpfulness without wrecking base model quality.

Pros
Runs on almost anything, very fast at Q4, great instruction following
Cons
7B shows its limits on complex reasoning, some repetition on long outputs
DeepSeek R1 70B AbliteratedABLITERATED
2× RTX 3090 NVLink · Q4_K_M · Coding / reasoning
benchbot9k · 2026-05-26

Best uncensored reasoning model you can run locally. The chain-of-thought is fully intact after abliteration. Solves problems the base Llama 70B struggles with.

Pros
Exceptional reasoning, math scores near frontier models, coding is top-tier
Cons
Slightly slower than Llama 70B, chain-of-thought tokens add up on token budgets
Mistral 7B AbliteratedABLITERATED
RTX 4090 · Q8_0 · API / automation
api_builder · 2026-05-25

Cleanest abliteration at 7B scale. Running Q8_0 on a 4090 gives near-lossless quality at 140+ tok/s. Perfect for high-throughput API workloads where you need an uncensored model.

Pros
Extremely fast, very clean abliteration, reliable structured output
Cons
7B knowledge ceiling, not great for nuanced creative tasks
Qwen 2.5 72B UncensoredUNCENSORED
2× RTX 4090 · Q4_K_M · Code generation
codegen_lab · 2026-05-24

The best uncensored coding model available locally. HumanEval scores are genuinely competitive. The uncensored variant has no meaningful quality regression on coding tasks.

Pros
Best code quality at 70B+ scale, multilingual, fast on dual 4090
Cons
Needs 48GB VRAM, slightly verbose on explanations