# DeepSeek V4.1 Flash is now available on GraphN

_DeepSeek's latest open-weights model is live as managed inference with a 1,048,576-token context window, image input, a thinking toggle, and tool calling._

[DeepSeek V4.1 Flash](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017) is now available in GraphN under the model alias **`deepseek-v4.1-flash`**. Teams can select it in the model catalog, attach it to an agent, or call it through GraphN's OpenAI-compatible inference API.

GraphN serves the 1017 release, published by DeepSeek on October 17, 2026. It is a mixture-of-experts vision-language model with 763B total parameters, and its weights are released under the MIT license. GraphN supports its full **1,048,576-token context window**.

Pricing is **$0.30 per 1M input tokens** and **$1.20 per 1M output tokens**.

## What DeepSeek V4.1 Flash adds

DeepSeek describes the 1017 release as a minor update to DeepSeek-V4.1-Flash with improvements in reasoning, coding, agentic tasks, and visual benchmarks, and with shorter outputs. On GraphN it brings:

- **Text and image input** for screenshots, charts, diagrams, and scanned pages alongside text
- **Thinking mode** on by default, with request-level control to turn it off when a direct answer is preferable
- **Tool calling** for agents that need to act on external systems
- **1M context** for large codebases, long documents, and extended task histories
- **MIT-licensed open weights** for teams evaluating portability and deployment control

In the released checkpoint, the expert weights are stored in MXFP4, a 4-bit floating-point format, and a separate vision encoder turns each image into tokens the language model reads alongside the text.

## What the reported benchmarks show

DeepSeek's model card compares the 1017 release with the earlier DeepSeek-V4.1-Flash. These are **vendor-reported results**, not GraphN benchmarks, and DeepSeek ran some of them in specific agent harnesses. They are useful directional evidence, not a substitute for testing the model on your workload.

| Workload                | V4.1 Flash 1017 | V4.1 Flash |
| ----------------------- | --------------: | ---------: |
| Terminal-Bench 4.0      |            42.3 |       31.2 |
| DeepSWE v1.1            |            77.0 |       74.2 |
| Automation-Bench v1.0.6 |            60.5 |       54.8 |
| Apex Bench v1.1         |            72.1 |       64.6 |
| Chartography (w/tools)  |            83.0 |       78.9 |

The reported gains span terminal coding, repository work, automation, agentic tasks, and chart understanding. That makes DeepSeek V4.1 Flash a candidate for agents that combine long context, tool use, and visual input in the same task.

## DeepSeek V4.1 Flash on GraphN

GraphN exposes the model through the same surfaces as its other built-in chat and vision models:

1. Open **Models** in a GraphN workspace.
2. Select **DeepSeek V4.1 Flash**.
3. Choose the model on an agent or workflow node.
4. On an agent, set **Enable Thinking** to **Off** for fast, direct answers. **Default** and **On** both leave thinking on for this model, which suits harder tasks.

The public alias is `deepseek-v4.1-flash`. The underlying checkpoint remains `deepseek-ai/DeepSeek-V4.1-Flash-1017`, so prompts and model behavior stay aligned with the released weights.

You can also test it from the GraphN CLI. Thinking is on by default, and the reasoning counts against the output budget, so allow a few thousand output tokens:

```bash
graphn model chat deepseek-v4.1-flash \
  --max-tokens 8192 \
  -m "Review this implementation plan, identify the highest-risk assumption, and propose a test."
```

For applications, GraphN keeps the inference boundary OpenAI-compatible. Existing clients can switch the `model` field to `deepseek-v4.1-flash` while retaining the standard chat-completions request shape, including image content in user messages.

To turn thinking off for a request, send `enable_thinking: false` explicitly. Leaving the field out keeps thinking on:

```json
{
  "model": "deepseek-v4.1-flash",
  "max_tokens": 512,
  "chat_template_kwargs": { "enable_thinking": false },
  "messages": [{ "role": "user", "content": "Summarize this changelog entry in one sentence." }]
}
```

In our own measurement on October 18, 2026, a single request generated about 138 tokens per second, with about 0.5 seconds to the first token. Those are single-request figures, not throughput or latency guarantees; measure your own prompts and concurrency before setting targets.

## Thinking and output budgets

Thinking is on unless you turn it off, and the reasoning counts against `max_tokens`. With a small budget, the whole budget can go to reasoning, and the response returns empty content with `finish_reason` set to `"length"`. Turn thinking off, or allow a few thousand output tokens, whenever you set a limit.

This alias accepts text and image input only; it does not accept video or audio. For video understanding, [Qwen3.8-27B](/blog/qwen3-8-27b-api-graphn) accepts video input on GraphN.

## When to choose it

Choose DeepSeek V4.1 Flash when the same agent needs several of these at once:

- coding and agentic work over large codebases or long histories
- questions over long documents, screenshots, and charts
- a context window of up to one million tokens
- controllable thinking for a mix of fast and difficult requests
- an open-weight model with a permissive license

As with any model change, evaluate it against representative tasks, latency targets, and failure cases before moving a production workflow.

## Start with a real task

The fastest useful evaluation is not a generic chat prompt. Give DeepSeek V4.1 Flash one task that combines the capabilities you expect to use in production: a repository plus an issue, a long document plus a structured-output contract, or a set of screenshots plus a tool.

Open your GraphN workspace and select `deepseek-v4.1-flash`, or review the [models guide](/docs/models) to begin.

## Primary sources

- [DeepSeek-V4.1-Flash-1017 model card](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017)
- [DeepSeek-V4.1-Flash-1017 configuration](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017/blob/main/config.json)
- [DeepSeek-V4.1-Flash-1017 license](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017/blob/main/LICENSE)
- [DeepSeek-V4.1-Flash model card](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash)
- [GraphN models reference](https://graphn.ai/docs/models)