Docs · Models

GLM-5.3-Abliterated

2026-10-24

A fine-tuned GLM 5.3 on a private deployment: text only, 1M-token context, three reasoning modes, priced above list, and covered by a compliance notice.

GLM-5.3-Abliterated is a commercially fine-tuned variant of Z.ai's GLM-5.3, served from a private deployment rather than the provider's official API. The tuning adds vertical business skills and removes the built-in refusal alignment (that is what "abliterated" means). It is labelled as a private deployment. The catalogue also includes deepseek-flash-Abliterated, a separate private model that accepts text and images.

The model id is case-sensitive. Send it exactly as GLM-5.3-Abliterated. An id in the wrong case is not in the catalogue, and the relay answers 403 with the message "Model is not offered by this service".

Route and endpoint

The model is sold on one route only, shown on the API keys page as GLM with the line name private deployment (route id z-official). There is no shared pool behind it, so a request on this route goes only to its dedicated upstream. An Auto route key dispatches the model to the same place.

Send requests to POST /v1/chat/completions:

curl https://ai.topxea.com/v1/chat/completions \
  -H "content-type: application/json" \
  -H "Authorization: Bearer sk-xxxx" \
  -d '{"model":"GLM-5.3-Abliterated","messages":[{"role":"user","content":"Say hello in one sentence."}]}'

The model field of the response says GLM-5.3-Abliterated, not the upstream's internal id. The same holds when you stream: every chunk, and the final chunk that carries usage, names the catalogue model.

/v1/messages also answers for this model: the relay converts the request to Chat Completions, with the attachment rewrite described below.

Text only

This deployment takes text and tool calls only. If a message carries an image, audio, file or video part, the relay does not fail the request with a 400. It replaces each such part with one text part that reads:

[Attachment omitted: this model accepts text only.]

The rest of the message is forwarded as sent, and the model sees the note in its context. Some clients send a small test image with their "add model" probe; with this rewrite the probe still passes. The rewrite applies on /v1/chat/completions and on /v1/messages, where it also covers images that a tool_result carried. A document inside a tool_result is still a 400.

The relay also drops, for every model but relevant here because this upstream rejects them, a tool_choice with no tools and empty containers such as "tools": [], "logit_bias": {} or "stop": []. Real tools, biases and stop sequences are forwarded untouched.

Context and output

The context window is 1,000,000 tokens, prompt and output together. /api/pricing publishes it as context_window and gives 131,072 as max_output, the output limit of Z.ai's own GLM-5.3. That figure is for clients that reserve room for the answer inside the window; the relay does not cap requests with it. The deployment accepts max_tokens or max_completion_tokens up to 999,990, though an answer that long only fits after a short prompt, and rejects 1,000,000 or more with a 400. A request without a cap is forwarded without one.

Reasoning

The model reasons before every answer. Set the depth with reasoning_effort. The deployment has three modes, low, high and max, and maps the standard values onto them: none, minimal and low run as low, medium and high as high, xhigh and max as max. A request that sets no effort runs as max. The nested form "reasoning": {"effort": "low"} works the same way.

Reasoning cannot be switched off. none runs as low and leaves the reasoning text out of the response; the tokens are still spent.

Z.ai's own API switches reasoning with "thinking": {"type": "enabled"} or {"type": "disabled"}, and this deployment rejects that object with a 400. The relay rewrites it before forwarding: disabled, and the boolean "thinking": false, become reasoning_effort: "none"; enabled is dropped because reasoning is on anyway. If the request also sets reasoning_effort or a reasoning object, that setting decides and the switch is dropped.

The reasoning text arrives in choices[].message.reasoning, or choices[].delta.reasoning when you stream, not in reasoning_content. Its tokens are counted in usage.completion_tokens, listed on their own as usage.completion_tokens_details.reasoning_tokens, and billed at the output price. To leave the text out of the response, send "reasoning": {"effort": "high", "exclude": true}: the model still reasons and the tokens are still billed. include_reasoning is not forwarded.

Structured output

A response_format of type json_schema works, with strict true or false: the deployment holds the answer to the schema while it writes it. That enforcement goes wrong for some string formats. In our tests (2026-10-16, reasoning_effort low), when the prompt did not say what a field should contain, a field with "format": "uri" filled up with run-together letters until the answer hit max_tokens, so the JSON came back cut off and every token up to the cap was billed; this happened in 9 of 12 tries, and "format": "email" did the same in 1 of 10. Fields with "format": "date" or "date-time" came back in the right shape with the wrong date: asked for Ada Lovelace's birth date, the model wrote 1046-10-14 instead of 1815-12-10, and 7 of 10 tries went wrong this way.

So before forwarding, the relay takes uri, email, date and date-time out of the schema and writes each into its field's description as {format: "uri"}. The deployment then only holds the field to a string. Prompts that asked for the value got a correct URL, address or ISO date 16 times out of 16, with or without the format. Prompts that left the field open often got a sentence of prose there instead; 2 of 40 such answers still ran to max_tokens and 1 had a wrong ISO date, while the same prompts with no format at all ran to max_tokens in 1 of 16. Other formats are forwarded as written. The two we also tried with open prompts, hostname and uri-reference, came back in the right shape with meaningless content (a hostname of 8, a uri-reference of Ada_Lovelace). Say in the prompt what each field should hold, and check the values in your own code after parsing.

Price

Per 1M tokens on the private deployment: input $4.00, output $7.00, cache read $0.40. That is above the Z.ai list price of the base GLM-5.3 (input $1.40, output $4.40, cache read $0.26, verified 2026-09-18). On the home page price list that reference row is labelled Provider list price · GLM-5.3, with the percentage above list next to each lane; the source is https://docs.z.ai/guides/overview/pricing.

Compliance notice

This notice appears in small print wherever the model is described, and it applies to every request:

Because fine-tuning may affect safety alignment, the model may produce biased, offensive or locally unlawful content. Users bear full responsibility for compliance and legal consequences.

The AI Acceptable Use Policy covers this line exactly as it covers Claude, OpenAI, Grok, Kimi, DeepSeek and Jev. Removing the refusal alignment does not remove the policy. If you are unsure whether a use case is allowed, ask through the help button at the bottom right or support@topxea.com before building on it.

Models: GLM-5.3-Abliterated · Models & pricing

Previous: Jev (TypeSafe System One) · Next: Image generation

Other languages: 简体中文