Z.aiZ.ai·💬 Text Generation·New

GLM 5.3 Flash Uncensored

ReasoningVisionCodeFunction CallingWeb Searchfp8private
🧠 Try in Intelligence →Try on Venice.ai ↗
Quick reference
GLM 5.3 Flash Uncensored — TLDR
  • 🔒 Abliterated build of GLM-5.3 Flash with refusal behavior removed from weights
  • 🧠 Per the publishers' model cards, a mixture-of-experts reasoning model, 320B total and 18B active
  • 📏 1M-token context window, described as retained unchanged from the base checkpoint
  • 👁️ Native vision tower preserved; accepts images alongside text
  • 🔧 Tool calling, structured output, and speculative decoding head still functional
  • ⚡ Shipped in block-FP8 according to the publishers' model cards
  • 📚 MIT-licensed lineage; community quantizations exist down to ~2.5 bits per weight
  • 🎯 No safety layer present; operators must supply their own
💰 Pricing
$0.165 / $0.550
per 1M · input / output
📏 Context
1.0M tokens
📅 On Venice since
Nov 13, 2026
4 days ago
Provider

Z.ai, formally Knowledge Atlas Technology Joint Stock Co., Ltd., is a Chinese technology company specializing in artificial intelligence. Previously known internationally as Zhipu AI, the company rebranded to Z.ai in 2025. Its core focus is the GLM family of…

Read full profile →
16 models on Venice
15 text · 1 image
Since Apr 1, 2024

About this model

GLM 5.3 Flash Uncensored is a community-modified derivative of Z.ai's GLM-5.3 Flash, the compact multimodal member of the GLM 5 series. It is not a new pretraining run: the publishers describe it as an abliterated checkpoint, produced by orthogonalizing the refusal direction out of the residual stream so that guardrails are removed at the weight level rather than through adapters or runtime filtering. The publishers' model cards describe the architecture as unchanged from the base model — a 320B-parameter mixture of experts with roughly 18B active parameters, hybrid linear and sparse attention, a native vision and video tower, a multi-token-prediction speculative head, and a one-million-token context.

Compared with its direct parent GLM 5.3 Flash, the practical difference is behavioral rather than architectural: the model answers prompts the aligned checkpoint declines. The publishers note guardrail removal is strongest in default and maximum reasoning-effort modes, and that very small output budgets can starve the reasoning phase in multi-turn chats.

Relative to earlier family generations such as GLM 5.2 and GLM 5, the Flash line is natively multimodal and uses the sparse/linear hybrid attention associated with its million-token window.

Because this is an unaligned research build, no content-filtering layer is present in the weights, and any moderation must be added by the operator. It is distributed in FP8 and in community quantizations spanning roughly 2.5 to 4 bits per weight.

This About section is AI-generated from public sources (Claude Opus 5), with no human editing. It may contain inaccuracies — verify critical details against the sources listed above.

Data sources: Venice API · HuggingFace · Wikipedia — enrichment updated 1d ago