Docs · Models

GLM-5.3-Abliterated and GLM-5.3-Flash-Abliterated

2026-11-10

The two GLM models on a private deployment: text-only GLM-5.3-Abliterated and the faster GLM-5.3-Flash-Abliterated with images. Reasoning, limits, prices.

The GLM line sells two models. Both are served from a private deployment rather than Z.ai's official API, and both are labelled private deployment wherever they appear.

• GLM-5.3-Abliterated is a commercially fine-tuned variant of Z.ai's GLM-5.3. The tuning adds vertical business skills and removes the built-in refusal alignment (that is what "abliterated" means).
• GLM-5.3-Flash-Abliterated is Z.ai's GLM-5.3-Flash with the refusal alignment removed the same way. It is cheaper and faster, and it accepts images.

The catalogue's third abliterated model, deepseek-flash-Abliterated, is on the DeepSeek line.

The weights row describes the private deployment: it runs quantized weights, not Z.ai's own full-precision service.

Model ids are case-sensitive. Send them exactly as GLM-5.3-Abliterated and GLM-5.3-Flash-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 endpoints

Both models are 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 them. An Auto route key dispatches them to the same place.

Send requests to POST /v1/chat/completions:

curl https://api.topxai.xyz/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 names the catalogue model, 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 both models: the relay converts the request to Chat Completions, with the attachment rewrite described below for GLM-5.3-Abliterated. GLM-5.3-Flash-Abliterated also takes POST /v1/responses; send the full conversation in input, because server-side conversation storage is not promised. GLM-5.3-Abliterated has no Responses endpoint.

Images

GLM-5.3-Abliterated 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.

GLM-5.3-Flash-Abliterated reads images sent as image_url parts with a base64 data URL (data:image/png;base64,...). It does not fetch http(s) image URLs: with a URL, the model answers that it cannot see the image, and the request is still billed.

The relay also drops, for every model, 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 on GLM-5.3-Abliterated and 1,048,576 on GLM-5.3-Flash-Abliterated, prompt and output together. /api/pricing publishes it as context_window and gives 131,072 as max_output for both. That figure is for clients that reserve room for the answer inside the window; the relay does not cap requests with it, and a request without a cap is forwarded without one. The deployment rejects a max_tokens or max_completion_tokens above 1,000,000, and a long cap only fits after a short prompt.

Reasoning

Both models reason before every answer, and reasoning cannot be switched off. Set the depth with reasoning_effort: low, high or max. A request that sets no effort runs as low. Until November 8, 2026 GLM-5.3-Abliterated defaulted to max; send "reasoning_effort": "max" to keep that depth.

The relay fits the other OpenAI values onto the three modes: minimal runs as low, medium as high and xhigh as max. none and off also run as low, because the models cannot stop reasoning, and the reasoning tokens are still spent. The nested form "reasoning": {"effort": "high"} works the same way, and on /v1/responses the effort goes in reasoning.effort. The deployment receives the effort only, so other members of the reasoning object, such as exclude, have no effect.

Z.ai's own API switches reasoning with "thinking": {"type": "enabled"} or {"type": "disabled"}. The relay folds that object into the effort before forwarding: disabled, and the boolean "thinking": false, become none, which runs as low; enabled is dropped because reasoning is on anyway. If the request also sets reasoning_effort or a reasoning object, that setting decides.

The reasoning text arrives next to the answer in choices[].message, or choices[].delta when you stream, as reasoning or reasoning_content. Its tokens are counted in usage.completion_tokens and billed at the output price. include_reasoning is not forwarded.

Parameters

Both models honour stop, response_format of type json_object or json_schema (strict or not), tool_choice set to "none", "required" or a named function, and parallel_tool_calls: false. Tools, parallel tool calls, tool-result round trips, streaming with usage, temperature and top_p work as in the OpenAI format. Two fields do not:

• n is not supported. One choice is returned and billed once, whatever n asks for.
• logprobs are not returned.

Structured output on GLM-5.3-Abliterated

On GLM-5.3-Abliterated the deployment holds the answer to a json_schema while it writes it, and that enforcement went 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 a request for GLM-5.3-Abliterated, 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. GLM-5.3-Flash-Abliterated receives the schema as sent.

Price

Prices in USD per 1M tokens on the main deployment, fixed at all hours, with no long-context tier and no cache-write lane:

These prices apply from November 10, 2026; the changelog lists the earlier ones. Prompt caching is automatic: cached prompt tokens are billed at the cache-read price. The pricing pages also show Z.ai's list price of each base model for reference, from https://docs.z.ai/guides/overview/pricing (checked 2026-11-08): GLM-5.3 at $1.40 input, $0.26 cache read and $4.40 output; GLM-5.3-Flash at $0.15, $0.03 and $0.50. Both models are above that list on every lane. Accepted requests and historical charges keep the prices captured when they were accepted.

Backup deployment

GLM-5.3-Abliterated runs on two private deployments. The second is a costlier standby. If the main deployment fails before it answers, the request may move to the standby automatically, and that request is billed at the backup price: $4.00 input, $0.40 cache read and $7.00 output per 1M tokens. This is rare.

You can recognise such a request in the usage log: its entry is marked Backup deployment and shows the backup rates it was billed at. The price tables on the home page and the model page list the same rates in a Backup deployment row under the model's price.

GLM-5.3-Flash-Abliterated has no standby. If its deployment fails, the request fails and nothing is charged.

Compliance notice

This notice appears in small print wherever either 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@topxai.xyz before building on it.

Models: GLM-5.3-Abliterated, GLM-5.3-Flash-Abliterated · Models & pricing

Related

• GLM-5.3-Abliterated API pricing
• GLM-5.3-Flash-Abliterated API pricing
• Acceptable use and abuse reports
• Shared pool, official line and auto route
• TopxAI vs the providers' own APIs: price, retention and what stays direct

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

Other languages: 简体中文