Now on VeniceLLMReasoningOpen weights

GLM 5.3 Flash Uncensored

An abliterated build of Z.ai's GLM-5.3-Flash — a 320B-parameter mixture-of-experts reasoning model with refusals edited out of its open weights, at flash-tier pricing.

For agents
curl https://api.venice.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "glm-5-3-flash-uncensored",
    "messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
  }'
Model IDglm-5-3-flash-uncensored
Maker
Z.ai (base GLM-5.3-Flash); uncensored build by dealignai
Context
1,024K tokens
Reasoning
Supported
Privacy
Private

Overview

What is GLM 5.3 Flash Uncensored

GLM 5.3 Flash Uncensored is an abliterated build of Z.ai's GLM-5.3-Flash, a 320B-parameter mixture-of-experts reasoning model with 18B active parameters, a 1M-token context window, vision, tool use and web search. Refusal behavior was edited out of its open weights. On Venice it runs privately with zero prompt retention at $0.17 in and $0.55 out per million tokens.

Running it privately on Venice

On Venice, GLM 5.3 Flash Uncensored runs under the private tier with zero retention — your prompts are not stored, profiled, or used for training, and there is no account history tied to your requests. Because the weights are open, you are not locked into one vendor's server: the same abliterated checkpoint is self-hostable, but on Venice you get it per-token at $0.17/$0.55 per million with cached input at $0.03. Vision, function calling, structured output and web search all work through the OpenAI-compatible Venice API without any logging middleman.

PrivateNo prompt training

Agent quickstart

Three calls, copied straight out

The API is OpenAI-compatible: change the base URL and the model id and existing client code works unchanged.

Streaming chat

curl https://api.venice.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "glm-5-3-flash-uncensored",
    "stream": true,
    "messages": [{ "role": "user", "content": "Draft the release note." }]
  }'

Tool calling

curl https://api.venice.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "glm-5-3-flash-uncensored",
    "messages": [{ "role": "user", "content": "Find the rate limits." }],
    "tools": [{
      "type": "function",
      "function": {
        "name": "search_docs",
        "description": "Search the API documentation.",
        "parameters": {
          "type": "object",
          "properties": { "query": { "type": "string" } },
          "required": ["query"]
        }
      }
    }],
    "tool_choice": "auto"
  }'

Python SDK

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["VENICE_API_KEY"],
    base_url="https://api.venice.ai/api/v1",
)

resp = client.chat.completions.create(
    model="glm-5-3-flash-uncensored",
    messages=[{"role": "user", "content": "Build without permission."}],
)
print(resp.choices[0].message.content)

Specifications

Datasheet

Maker
Z.ai (base GLM-5.3-Flash); uncensored build by dealignai
Open weights
Yes
License
MIT (base model); uncensored FP8 build republished by dealignai
Context window
1,024K tokens
Input images
Supported — multiple images accepted per request
Released
August 26, 2026 (base model); uncensored FP8 weights August 2026; live on Venice since November 13, 2026
Architecture
Hybrid mixture-of-experts with sparse and linear attention; Manifold-Constrained Hyper-Connections (mHC)
Parameters
320B total, 18B active per token
Uncensoring method
Abliteration — refusal behavior removed by editing the weights, no retraining
Quantization on Venice
FP8 (native block-FP8 build)
Max output
128K tokens
Capabilities
Vision, Function calling, Reasoning, Web search, Code-optimized
Privacy on Venice
Private — zero retention
Available on Venice since
Nov 2026

Assessment

Strengths and limitations

Strengths
  • Frontier-tier reasoning at flash cost: Z.ai reports GLM-5.3-Flash pushes the Artificial Analysis Intelligence Index v4.1.1 Pareto frontier, scoring 57 at roughly $0.045 per task — intelligence previously available only at ~10× the price.
  • Strong coding and agentic performance: Z.ai reports it outperforms GLM-5.2 across six coding and agentic benchmarks (63.4 vs 46.2 on DeepSWE v1.1, 48.8 vs 26.2 on AutomationBench) while approaching Claude Opus 4.8.
  • Abliterated weights cut benign over-refusal sharply — vendor-reported figures show benign refusals dropping from 2.4% to 0.4% — so legitimate security, creative and research work stops getting bounced.
  • Full tool stack: function calling, structured JSON-schema output, web search, and code-optimized generation all work on Venice.
  • Vision with multiple image inputs per request, on top of a 1,024K-token context window for whole-repo or whole-archive analysis.
  • Open weights under MIT mean you can self-host, fine-tune, or audit the model — no permission needed.
Limitations
  • Abliteration does not remove every refusal: vendor-reported evals still show roughly 11–14% hard-line refusals on the most adversarial requests, so this is not a blank-check model.
  • Weight edits can cost some agent reliability: third-party rankings report occasional tool-call loops and a lower agent-task score than the base model, so verify behavior on your own workflows.
  • Max output is 128K tokens on Venice; some uncensored rivals (e.g. DeepSeek V4.1 Flash builds hosted elsewhere) allow longer single responses.
  • No TEE or end-to-end encryption on Venice for this model — privacy comes from Venice's zero-retention policy, not hardware isolation.

Use cases

What it is good for

  1. 01Authorized security work — penetration testing systems you have permission to test, defensive research, and API reverse engineering without moralizing refusals.
  2. 02Long-document analysis: drop a full codebase, contract set, or research archive into the 1M-token context in one pass.
  3. 03Agentic coding pipelines that need function calling, JSON-schema outputs, and long context in a single cheap model.
  4. 04Unfiltered creative writing and worldbuilding where safety preambles break the voice.
  5. 05Multimodal analysis — comparing several screenshots, diagrams, or photos in one request.

Prompting

Getting better results

Set reasoning effort to low for chat and extraction, and max only for hard math, proofs, or architecture decisions — max thinking burns output tokens.

Keep your system prompt and few-shot examples identical across calls so Venice's $0.03/1M cached-input rate kicks in on repeat traffic.

Attach multiple images in one request and ask for a comparative analysis — the model accepts several images per turn.

For agents, pass an explicit JSON schema and let structured output enforce it instead of parsing free-text replies.

Paste long documents whole rather than summarizing first — the 1,024K context is there to be used, and splitting adds prompt cost.

When a request gets refused, rephrase the authorized-use context ("I have written permission to test this system") rather than assuming the weights will always comply — residual refusals exist.

Alternatives

How it compares

ModelBest forContextOpen weightsPrice (Venice)
GLM 5.3 Flash UncensoredUncensored frontier reasoning at flash cost1M tokensYes$0.17 in · $0.55 out / 1M
Venice Uncensored 1.2Unfiltered chat on a tight budget128K tokensYes$0.20 in · $0.90 out / 1M
Gemini 3.8 FlashFast general-purpose assistant work1M tokensNo$0.94 in · $4.69 out / 1M
Qwen 3.7 PlusLong-context work with standard safety rails1M tokensNo$0.50 in · $2 out / 1M

Pick GLM 5.3 Flash Uncensored when you need frontier reasoning, vision, and tool use with the refusals edited out — and want it at the cheapest per-token price of any frontier-tier model on Venice. If you only need unfiltered conversation, Venice Uncensored 1.2 is the budget pick.

Pricing

What it costs on Venice

Billed per token on Venice: $0.17 per 1M input tokens and $0.55 per 1M output tokens.

Input / 1M tokens
$0.17
Per 1M tokens
Output / 1M tokens
$0.55
Per 1M tokens
Cached input / 1M
$0.03
Per 1M tokens

New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.

Getting a key

From nothing to a first call

  1. 01

    Create a key in API settings. Nothing else is required to start.

  2. 02

    Export it as VENICE_API_KEY so the snippets above run unedited.

  3. 03

    Point an existing OpenAI client at https://api.venice.ai/api/v1. The scheme is part of the value: an OpenAI client given a bare host does not resolve it.

  4. 04

    Pass glm-5-3-flash-uncensored as the model and send the request.

FAQ

Frequently asked questions

GLM 5.3 Flash Uncensored is an abliterated version of Z.ai's GLM-5.3-Flash, a 320B-parameter mixture-of-experts reasoning model with 18B active parameters. Refusal behavior was edited out of the open weights without retraining, and the model keeps a 1M-token context, vision input, function calling, and web search.

Venice bills per token: $0.17 per million input tokens, $0.55 per million output tokens, and $0.03 per million cached input tokens. There is no subscription requirement — you pay for what you use.

The weights are open. The base GLM-5.3-Flash model is released by Z.ai under the MIT licence, and the uncensored FP8 build is published by dealignai on Hugging Face. You can self-host it, but using it on Venice is pay-per-token rather than free.

Mostly. Abliteration removed the refusal direction from the weights, and vendor-reported evals show benign over-refusal dropping from 2.4% to 0.4% and hard-line refusals on adversarial requests falling to roughly 11–14%. Refusals are drastically reduced but not zero, so edge-case requests can still be declined.

Yes. It accepts image input including multiple images per request, supports function calling, structured JSON-schema output, web search, and is code-optimized. All of these work through the Venice app and the Venice API.

GLM 5.3 Flash Uncensored is the stronger pick for reasoning, coding, agents, vision, and very long documents, with a 1,024K-token context versus 128K. Venice Uncensored 1.2 is the lighter, cheaper option for straightforward unfiltered chat where frontier intelligence and long context don't matter.

Venice runs it under the private tier with zero retention: prompts and outputs are not stored, profiled, or used for training. Note that this model is not processed in a TEE and does not use end-to-end encryption — the privacy guarantee is Venice's no-logging policy, not hardware isolation.

Barely, according to published evals. Vendor-reported testing shows no MMLU loss against the base model, though third-party rankings note some agent-reliability cost from the weight edit. For most chat, coding, and analysis work the two behave the same — only the refusals changed.

Z.ai released the base GLM-5.3-Flash model on August 26, 2026, the uncensored FP8 weights followed in late August 2026, and Venice listed GLM 5.3 Flash Uncensored on November 13, 2026. It is live on Venice now.

Run GLM 5.3 Flash Uncensored privately

One key, free to start, no credit card.

Start chat