DeepSeek V4.1 Flash (1017): What Changed, API, Open Weights and How to Try It

DeepSeek V4.1 Flash 1017 (DeepSeek-V4.1-Flash-1017) is a minor update to DeepSeek's V4.1 Flash model, released on 17 October 2026. DeepSeek says it scores higher on reasoning, coding, agentic and visual benchmarks while finishing tasks faster with shorter outputs, which lowers cost. It now powers the deepseek-flash model on DeepSeek's API, website and app, and the weights are on Hugging Face under the MIT licence. API pricing starts at $0.15 per million input tokens and $0.60 per million output tokens off-peak (as of October 2026).

Key facts

What changed in the 1017 release

DeepSeek's release note calls 1017 a Pareto improvement: "Stronger benchmark performance, faster task completion, and shorter outputs." It says the model was trained with reinforcement learning with self-feedback to improve end-to-end project development, and it's optimised for DeepSeek Harness (v0.2.1-rc.1 or later), Tencent CodeBuddy and WorkBuddy.

The API side changed too:

• deepseek-flash is now served by DeepSeek-V4.1-Flash-1017; calls don't need to change.
• The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are retired. They're still accepted, but route to 1017 and bill at the Flash price.
• DeepSeek's Search and Fetch APIs are open to everyone at $0.75 per 1,000 successful calls each. Calls made through DeepSeek Harness, Claude Code or Codex are free until 31 October 2026, 16:00 UTC.

Benchmarks DeepSeek published

From the 1017 model card (DeepSeek-reported; competitor columns are older models):

The card says Terminal-Bench 4.0 was run with the Claude Code harness and DeepSWE v1.1 with Mini-SWE-V2. The jump over the previous V4.1 Flash is large for a "minor update", especially on Terminal-Bench (+11.1 points) and ProgramBench (+8.8). The comparison set doesn't include the newest closed models such as GPT-6 or Gemini 4 Argon, so don't read it as a frontier ranking.

Architecture in brief

The 1017 card doesn't restate the architecture, so this comes from the V4.1-Flash card: a mixture-of-experts model with a causal encoder-decoder design (20 encoder and 20 decoder layers), 1 shared and 384 routed experts per MoE layer with 6 active per token, sparse attention, FP4 KV caching and native image input. DeepSeek says the global KV cache is about a quarter of V4-Flash's, which is why a 1M context is cheap to serve. If MoE is new to you, see mixture of experts.

Reasoning effort on V4.1 Flash is an integer from 1 to 100 on the model itself; the API takes thinking (the default) and non-thinking modes.

How to try DeepSeek V4.1 Flash 1017

1. In the browser or app

DeepSeek says 1017 is live on its website and app. That's the fastest way to get a feel for it. Some organisations restrict DeepSeek's consumer app; check your own policy before pasting work data.

2. Through the API

The API is OpenAI-compatible at https://api.deepseek.com, and Anthropic-compatible at https://api.deepseek.com/anthropic. DeepSeek's quickstart:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ.get("DEEPSEEK_API_KEY"),
    base_url="https://api.deepseek.com",
)

response = client.chat.completions.create(
    model="deepseek-flash",
    messages=[
        {"role": "system", "content": "You are a helpful assistant"},
        {"role": "user", "content": "Hello"},
    ],
    reasoning_effort="high",
    extra_body={"thinking": {"type": "enabled"}},
)
print(response.choices[0].message.content)

deepseek-flash supports JSON output, tool calls, the Responses API, the Anthropic API format, vision, and FIM completion (beta, non-thinking mode only), with a concurrency limit of 2,500. The Anthropic-format endpoint means you can point tools built for Claude at it; that's how DeepSeek's free Search/Fetch offer through Claude Code works.

3. Self-host the open weights

The model card's vLLM and SGLang commands:

pip install vllm
vllm serve "deepseek-ai/DeepSeek-V4.1-Flash-1017"

DeepSeek recommends temperature 1.0, top_p 1.0, a 1M context and max_tokens of at least 256K. There's no Jinja chat template; prompts are built with the Python reference in the repo's encoding folder. At 763B parameters in FP8 and BF16, this is a multi-GPU server job, not a laptop one. The V4.1-Flash card says quantised builds work with llama.cpp, Ollama and LM Studio; for what's realistic at home, see local LLM hardware and best local LLM for coding.

Pricing (as of October 2026)

Per 1M tokens, from DeepSeek's Models and Pricing page:

Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays; everything else, including weekends, is off-peak at half price. For comparison, deepseek-v4-pro (still listed) is $0.66 / $1.98 off-peak, with no vision support. If your batch jobs can wait, schedule them off-peak.

Who should use it

V4.1 Flash 1017 is a strong candidate for cost-sensitive coding agents and high-volume automation, where DeepSeek's AutomationBench and DeepSWE numbers are competitive and the price is a fraction of the closed models. The MIT weights also make it one of the few top-tier open-weight models you can fine-tune and ship without a custom licence. For how it stacks up against the rest, see the best LLM guide.