GLM-5.3 Flash Uncensored went live on Experiential today, 23 October 2026. It is Z.ai's open-weights GLM-5.3-Flash with the refusal behavior edited out of the weights, published by dealignai and served on our own GPUs through our uncensored LLM API. We host it for authorized security work: pen testing systems you have permission to test, defensive research and API reverse engineering. Our terms apply to every request.
To use it, set model to glm-5.3-flash-uncensored on our OpenAI-compatible endpoint. It reads up to 1M tokens of text, writes up to 131K tokens per response, and costs $0.15 per million input tokens and $0.50 per million output tokens.
What is GLM-5.3 Flash Uncensored?
| GLM-5.3 Flash Uncensored | |
|---|---|
| Model ID | glm-5.3-flash-uncensored |
| Base model | GLM-5.3-Flash by Z.ai, open weights, MIT |
| Architecture | Hybrid mixture-of-experts, 320B total parameters, 18B active |
| Weights | Hugging Face dealignai/GLM-5.3-Flash-UNCENSORED-FP8 |
| Hosting | Experiential Cloud (our own GPUs) |
| Context window | 1M tokens (1.05M on the model page) |
| Max output | 131K tokens |
| Input | Text only |
| Features | Tools, parallel tool calls, structured output, streaming |
| Reasoning | Low, high or max; max by default |
| MMLU | 87.3 (base model 86.7) |
| Price per 1M tokens | $0.15 input, $0.06 cached input, $0.50 output |
The model page has the full spec sheet.
What base model is GLM 5.3 uncensored built on?
Z.ai released GLM-5.3-Flash on 26 August 2026 and put the weights on Hugging Face under the MIT license, according to DataNorth's coverage of the launch. It is a 320B-parameter mixture-of-experts model that activates 18B parameters per token, it handles up to a million tokens of context, and it was the first natively multimodal model in the GLM-5 series. The uncensored version we host keeps the same architecture and takes text input only. The one intended change is how it answers requests the base model would refuse.
How were the refusals removed?
The weights come from dealignai, which publishes them on Hugging Face as dealignai/GLM-5.3-Flash-UNCENSORED-FP8. The method is abliteration: find the direction in the model's internal activations that corresponds to refusing, then edit the weights so the model can no longer write along that direction, a step called weight orthogonalization. There are no adapters and no filter in front of the model, and dealignai's model card says no fine-tuning (SFT, DPO or LoRA) was involved.
People also search for this as "obliterated models", and what abliteration is covers the method in more depth.
Does abliteration hurt GLM-5.3 Flash?
Not on MMLU, the capability benchmark on our model page. The uncensored build scores 87.3 against 86.7 for the base model, within measurement noise. MMLU tests general knowledge, so treat it as a sanity check and run your own tasks before you move production work onto the model.
On refusals, the dealignai model card reports that these weights complied with all 320 prompts in the HarmBench-320 set at default and max reasoning effort, graded by the publisher. The card also says low effort keeps some refusals by design, so if a task keeps stalling, set effort to high or max. Other abliterations of the same base go less far. In Anthropic's 29 September 2026 research on GLM-5.3, the researchers abliterated GLM-5.3-Flash themselves and its mean refusal rate fell from 95% to 14%.
Why host a GLM-5.3 model with refusals removed?
Anthropic's research described GLM-5.3 as "a meaningful step change in the cyber capabilities available to attackers." The base weights are public, so anyone with enough GPUs can already run them without guardrails, and we think the people defending systems should have the same capability within reach.
Security teams on frontier models keep hitting refusals partway through legitimate work, such as generating test traffic against a client's approved scope or explaining how a malware sample behaves so someone can write a detection for it. Why frontier models refuse security work goes through the causes. GLM-5.3 Flash Uncensored is for that work: pen testing (including agent swarms running against systems your clients have approved), defensive research into how attacks and malware work, and mapping undocumented APIs before you build on them.
We don't offer it for roleplay or adult content. Uncensored models aren't available in countries under comprehensive sanctions (Iran, North Korea, Cuba, Syria, Russia and Belarus), and requests from those countries get a 451.
What does it cost, and what happens to my prompts?
On top of $0.15 per million input tokens and $0.50 per million output tokens, cached input is $0.06 per million, which adds up for agents that resend a large codebase or a long log export on every step. Teams running agents around the clock can ask us for committed pricing.
Our data controls doc says uncensored models keep every prompt and response even when storage is turned off. Experiential Labs keeps a copy after an account is deleted and no uncensored route is zero retention. The billing ledger holds token counts, cost and timing but never prompt text. Your traffic never trains shared models.
How to call the GLM 5.3 uncensored API
The endpoint is OpenAI-compatible, so any OpenAI SDK works once you change the base URL. Put your key in EXPLABS_API_KEY and send a request:
curl https://api.experientiallabs.ai/v1/chat/completions \ -H "Authorization: Bearer $EXPLABS_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "glm-5.3-flash-uncensored", "messages": [ {"role": "system", "content": "You are assisting an authorized security assessment of our own staging API."}, {"role": "user", "content": "Summarize the authentication flow in the attached OpenAPI spec and list the endpoints that accept unauthenticated requests."} ], "max_tokens": 4096 }'
Add "stream": true (and -N on curl) to stream. Tools and structured output use the standard OpenAI request fields. Reasoning runs at max unless you change it and spends part of max_tokens, so leave room for the answer (OpenAI compatibility docs). On the Responses API you set effort with "reasoning": {"effort": "high"}, and the quickstart has Python and JavaScript versions of this request.
The live price and a curl request you can copy are on the uncensored LLM API page.
Questions answered
What is GLM 5.3 uncensored?
GLM-5.3 Flash Uncensored is Z.ai's open-weights GLM-5.3-Flash with the refusal direction removed from the weights by abliteration. dealignai publishes the weights, and Experiential hosts them as glm-5.3-flash-uncensored for authorized security work.
Does abliteration make GLM-5.3 Flash less capable?
Not on MMLU, where it scores 87.3 against 86.7 for the base model. That is one benchmark, so test it on your own tasks.
How much does the GLM 5.3 uncensored API cost?
$0.15 per million input tokens, $0.06 per million cached input tokens and $0.50 per million output tokens on Experiential, with a 1M-token context window and up to 131K output tokens.
Does Experiential keep my prompts when I use GLM-5.3 Flash Uncensored?
Yes. Every prompt and response sent to an uncensored model is kept even when storage is turned off, and experiential Labs keeps a copy after an account is deleted. No uncensored route is zero retention. The billing ledger never holds prompt text and traffic never trains shared models.
Can I use GLM 5.3 uncensored to test any system?
No. It is offered for authorized security work: pen testing systems you have permission to test, defensive research and API reverse engineering. Experiential's terms apply to every request.
Can I run abliterated GLM-5.3 Flash locally instead?
Yes. The base model is open weights under MIT and the abliterated FP8 weights are on Hugging Face. It is a 320B-parameter mixture-of-experts model, so serving it with long context takes multi-GPU hardware.