Docs · Models

deepseek-flash-Abliterated and deepseek-flash-fast

2026-11-12

The two DeepSeek V4.1 Flash 1017 models on a private deployment, abliterated or on standard weights: limits, four reasoning settings and fixed prices.

The DeepSeek line sells two models. Both are built on DeepSeek V4.1 Flash 1017 (DeepSeek's id DeepSeek-V4.1-Flash-1017), the current version, and both come from a private deployment rather than DeepSeek's official API.

• deepseek-flash-Abliterated is DeepSeek V4.1 Flash 1017 with the built-in refusal alignment removed (abliterated). It carries the compliance notice at the end of this page.
• deepseek-flash-fast is DeepSeek V4.1 Flash 1017 on its standard weights. It is not abliterated: it keeps the base model's refusal behaviour, carries no Abliterated notice, and is listed with the DeepSeek models, not in the Abliterated group.

Send the model IDs exactly as shown, including the capital A in Abliterated.

Route and limits

Both models are served on the existing deepseek-official route. That route ID remains for API keys, but neither model calls DeepSeek's official API; the API keys page shows the line as a private deployment with fixed prices. An Auto-route key can also reach them.

The weights row describes the private deployment: both models run FP8-quantized weights, not DeepSeek's own full-precision service.

Input and output must fit within the context window. The output limit holds even when more space remains in the context: 131,072 tokens on deepseek-flash-Abliterated, 262,144 on deepseek-flash-fast. If you omit the output cap, TopxAI fills it from that limit, reduced to the space remaining after the estimated prompt. This applies to max_tokens / max_completion_tokens on Chat Completions, converted Messages requests, and max_output_tokens on Responses.

Make a request

Use the existing TopxAI base URL and key:

curl https://api.topxai.xyz/v1/chat/completions \
  -H "Authorization: Bearer $TOPXAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-flash-Abliterated","max_tokens":4096,"reasoning_effort":"high","messages":[{"role":"user","content":"Explain how a database transaction works."}]}'

Put deepseek-flash-fast in model for the standard-weights model. Responses requests go to /v1/responses; Messages requests are converted to Chat Completions. Responses supports stateless requests: clients can send the full conversation context in input. TopxAI does not promise server-side conversation persistence for this deployment. Endpoint support alone does not establish that every Codex workflow has been tested.

Set stream: true for streaming. For an image request, put text and image_url content blocks in the user message as in the OpenAI Chat Completions format. deepseek-flash-fast reads only base64 data URLs (data:image/png;base64,...): it does not fetch an http(s) image URL, the model answers that it cannot see the image, and the request is still billed. deepseek-flash-Abliterated fetches image URLs as well. The API reference describes the Messages conversion and how images in tool results are handled.

Reasoning

Both models accept none, low, high and max in reasoning_effort (reasoning.effort on Responses). A request that sets no effort runs as low. none, or off, turns reasoning off. The relay fits the other OpenAI values: minimal runs as low, medium as high and xhigh as max. The "thinking": {"type": "disabled"} switch that DeepSeek documents becomes none; enabled is dropped. Until November 8, 2026 deepseek-flash-Abliterated defaulted to high; send "reasoning_effort": "high" to keep that depth.

Parameters

On both models, tools, parallel tool calls, tool-result round trips, streaming with usage, Responses streaming and tools, Messages streaming, temperature, top_p and a json_object response format work. Two fields do not, on either model:

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

deepseek-flash-Abliterated also honours stop, a strict json_schema response format, tool_choice set to "none" or "required", and parallel_tool_calls: false.

deepseek-flash-fast ignores those four: stop sequences, a json_schema response format (use json_object and check the result in your own code), tool_choice set to "none" or "required", and parallel_tool_calls: false. A tool_choice that names one function works. It may also write past a very small max_tokens.

Fixed prices

Default prices in USD per million tokens on the main deployment:

The deepseek-flash-Abliterated prices apply from November 10, 2026; the changelog lists the earlier ones. The same prices apply throughout the day and across the context window. There is no peak/off-peak tariff, no long-context premium and no cache-write lane. Prompt caching is automatic, and cache-read pricing applies to tokens reported as cache hits. See the pricing pages for current prices and the usage log for the token categories charged on a request. Accepted requests and historical charges keep their captured prices.

For reference, the pricing pages show DeepSeek's own list price for deepseek-flash, which DeepSeek serves with DeepSeek-V4.1-Flash-1017 (https://api-docs.deepseek.com/quick_start/pricing, checked 2026-11-08): $0.15 input, $0.003 cache read and $0.60 output off-peak, and $0.30, $0.006 and $1.20 at peak, which DeepSeek sets at 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday. Those hours do not apply to the two private-deployment models here.

Backup deployment

deepseek-flash-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: $1.30 input, $0.15 cache read and $2.60 output per million tokens (since 2026-11-12; $1.20 / $0.15 / $2.40 before). 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.

deepseek-flash-fast has no standby. If its deployment fails, the request fails and nothing is charged.

Replacing the old model

deepseek-flash is retired. Requests to the old ID receive HTTP 403; there is no forwarding alias. Update the model ID in saved clients and integrations. If your key has an allowed-model list, add deepseek-flash-Abliterated or deepseek-flash-fast separately. Restricted keys are not automatically granted access to either model.

Check saved output limits too: deepseek-flash-Abliterated writes at most 131,072 tokens, and a request copied from an older configuration may ask for more. The private deployment also has its own reasoning settings; do not assume that every setting from the official model transfers.

Compliance notice

This notice applies to deepseek-flash-Abliterated: 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.

deepseek-flash-fast carries no such notice. The Acceptable Use Policy applies to both models, as it does to the GLM models and the rest of the catalogue.

Models: deepseek-flash-Abliterated, deepseek-flash-fast · Models & pricing

Related

• deepseek-flash-Abliterated API pricing
• deepseek-flash-fast API pricing
• Moving to deepseek-flash-Abliterated on TopxAI
• Create an API key and choose a route
• How requests are billed
• Acceptable use and abuse reports

Previous: Video generation · Next: Kimi K3.1 Pro

Other languages: 简体中文