I Tested Every Uncensored Local AI Model (8GB to 32GB VRAM)

I Tested Every Uncensored Local AI Model (8GB to 32GB VRAM)

Title history
  • I Tested Every Uncensored Local AI Model (8GB to 32GB VRAM) — observed on 20261115-1639 UTC (YouTube)
🎙 Kai 👥 75K 📅 November 15, 2026 ⏱ 29 min 👁 2K 📄 tutorial 🧭 2026-11-15
Available in: English (current) Français

Keywords

uncensored modelsVRAMtoken budgetOllamarefusals

Summary

The video investigates uncensored local AI models for GPUs with 8-32 GB VRAM. The creator first faces absurd refusals (e.g., a Pac-Man clone) and discovers a vast ecosystem of ‘uncensored’ models on Hugging Face claiming zero refusals. However, the refusal count is misleading. A rigorous independent test by Nathan Sapwell on eight uncensored Qwen 3.8 27B builds (400 harmful requests, 167 GPU hours) reveals that refusals are essentially solved (best build refused only 2/400), but up to 38% of answers never arrive due to thinking loops exhausting the token budget. The token budget is tied to VRAM, and Ollama defaults to only 4,096 tokens under 24 GB, which is too small. The fix is setting OLLAMA_CONTEXT_LENGTH to at least 16,000. The video also warns that aggressive uncensoring degrades truthfulness (up to 14 points), and that download counts don’t indicate quality. It debunks fake model names and modified chat templates. Finally, it provides specific model recommendations for each VRAM tier, balancing finish rate, accuracy, and context size.

166 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video provides valuable, actionable insights by moving beyond marketing claims about refusal counts. It emphasizes the importance of finish rate and the practical limitations imposed by VRAM and context length. The argumentation is solid, grounded in Sapwell’s systematic benchmark and the creator’s own trials. It effectively explains why uncensored models sometimes produce blank responses and offers a simple fix. The reasoning is clear and avoids hype, though it relies heavily on a single external test.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is moderate: the main source is an independent benchmark with a well-defined methodology, but no direct links or peer-reviewed references are provided. The video also includes personal observations and identifies common pitfalls (e.g., misleading cards, template jailbreaks). The title is slightly ambitious but conveys the scope accurately. Overall, the content is trustworthy within the context of the single test and the creator’s practical experience.

158 words

Title / Content Match

The title broadly reflects the content, although 'every' is an exaggeration since the video focuses on a selection of models and a specific test. This overstatement does not distort the core message.

Quality & Reliability

7/10

The video relies on a rigorous independent test by Nathan Sapwell (400 harmful requests, 167 GPU hours) and extensive personal troubleshooting, providing concrete numbers and practical advice. However, it depends almost exclusively on a single external source, lacks direct citations, and some conclusions are based on anecdotal experience.

Chapters

Cited Sources

  • Nathan Sapwell's independent test on uncensored Qwen 3.8 27B builds — Referenced as the main benchmark: 400 harmful requests, 167 GPU hours, RTX 5090, 8 builds.
  • Harbench — Benchmark used in the test to measure refusals on 400 harmful requests.
  • Hugging Face — Platform where uncensored models are hosted and downloaded.
  • Ollama — Tool used to run local LLMs; context length default and OLLAMA_CONTEXT_LENGTH setting.
  • Heretic — Open-source tool used to create uncensored models via abliteration; mentioned with over 37,000 stars on GitHub.

Contribution & Novelties

The video’s original contribution is a clear, practical guide that reorients users from refusal counts to finish rates and context management. It exposes common pitfalls like fake model names and template jailbreaks, and provides tier-specific recommendations. It also bridges the gap between technical benchmarks and user experience.

Pour aller plus loin :

  • Chain-of-thought prompting — Relevant to understanding thinking models and why drafts consume tokens.
  • Quantization (signal processing) — Explains the trade-off between model size and quality, central to VRAM constraints.
  • Context window — Concept of token budget and its impact on model outputs.
  • Abliteration (technique) — No reliable URL available; mentioned as the method to remove refusals.

108 words

Radar Profile

The score profile indicates high information quantity and quality, with moderate technical depth and reliability. This suggests the video excels at delivering detailed, actionable content but might not go deep into underlying theory or verify with multiple sources.

Reliability 7/10