YOU ARE VIEWING ONE ITEM FROM THE AICRIER FEED

Nanbeige 4.3 33B Goes Local, Uncensored

AICrier tracks AI developer news across Product Hunt, GitHub, Hacker News, YouTube, X, arXiv, and more. This page keeps the article you opened front and center while giving you a path into the live feed.

// WHAT AICRIER DOES

7+

TRACKED FEEDS

24/7

SCRAPED FEED

Short summaries, external links, screenshots, relevance scoring, tags, and featured picks for AI builders.

Nanbeige 4.3 33B Goes Local, Uncensored
OPEN LINK ↗
// 1h agoMODEL RELEASE

Nanbeige 4.3 33B Goes Local, Uncensored

The dealignai community release is a refusal-removed GGUF build of Nanbeige4.3-33B, preserving its advertised 512K context, switchable thinking, and agentic coding focus. It ships multiple quantizations for local inference, with self-reported 100% HarmBench compliance that is not independent safety validation.

// ANALYSIS

This is less a new frontier model than a highly practical local distribution of Nanbeige’s newly released 33B model—and the uncensored angle is the real headline.

  • –Official Nanbeige evaluations beat Qwen3.8-27B on Terminal-Bench 2.1 and DeepSWE, but trail it on SWE-Bench Pro.
  • –The uncensored build reports 100% compliance across HarmBench categories, meaning it fulfills restricted prompts rather than demonstrating safety.
  • –Q4_K_M offers the best quality-to-size tradeoff at roughly 19GB; smaller IQ2 builds fit constrained hardware but lose more MMLU accuracy.
  • –MMLU remains close to the base model at Q4, while the lowest-bit quantization shows a larger quality drop.
  • –The upstream model is Apache 2.0, but this community variant uses a separate DealignAI license with restrictions on hosted inference.
// TAGS
nanbeige4.3-33b-uncensored-ggufllmopen-weightslong-contextquantizationlocal-firstinference

DISCOVERED

1h ago

2026-10-26

PUBLISHED

2h ago

2026-10-26

RELEVANCE

9/ 10

AUTHOR

0x0SojalSec