DeepSeek V4.1 Flash 1017¶

DeepSeek V4.1 Flash 1017 is DeepSeek's October 2026 update to DeepSeek V4.1 Flash. It is a Mixture-of-Experts model that reads text and images, works with up to 1M tokens of context, and can reason before it answers when you ask it to.

Model details¶

Prices and your rate limits are on the model's page in the Console. Image input works through the API; the Playground in the Console takes text only.

Good for¶

According to DeepSeek, the 1017 update improves on V4.1 Flash in reasoning, coding, agentic and visual tasks, and finishes tasks faster with shorter outputs. It suits:

• Coding agents: fixes and features in real repositories, and projects built from scratch
• Agentic workflows: tool use, office automation and multi-step tasks
• Image understanding: screenshots, charts and documents, with or without tools
• High-volume chat and extraction: low price per token, and cheaper cached input

Benchmarks¶

DeepSeek's published results for the 1017 update, next to the previous V4.1 Flash:

For the comparison with other models and the harnesses used, see the model card. Test it on your own tasks before you choose.

Call the model¶

curl -X POST "https://inference.verda.com/v1/chat/completions" \
  -H "Authorization: Bearer <your key>" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-4-1-flash-1017",
    "messages": [{"role": "user", "content": "What is the best viewpoint in Helsinki?"}],
    "max_tokens": 2048
  }'

For Python, Node and coding agents, see the Quickstart and Coding agents.

Reasoning¶

DeepSeek V4.1 Flash 1017 answers directly by default. To have it reason first, set reasoning_effort to low, high or max. It returns its reasoning in reasoning_content, and reasoning counts as output tokens. To turn reasoning off again, leave out reasoning_effort or set it to none.

DeepSeek's own API switches reasoning with a thinking parameter. Verda ignores it, so use reasoning_effort instead.

curl -X POST "https://inference.verda.com/v1/chat/completions" \
  -H "Authorization: Bearer <your key>" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-4-1-flash-1017",
    "messages": [{"role": "user", "content": "How many weekdays are there in March 2027?"}],
    "reasoning_effort": "high",
    "max_tokens": 16384
  }'

With reasoning on, leave room in max_tokens so the answer isn't cut off after the reasoning.

When the request includes tools and reasoning is on, send back each earlier assistant message in full, with its reasoning_content and tool_calls, not only content. If you leave out reasoning_content, the model sees an empty reasoning block where its earlier reasoning was. Without tools, earlier reasoning is left out of the prompt, so you don't need to send it back.

Cached input tokens cost less than regular input tokens, so long prompts that repeat the same opening are cheaper to send again.