DeepSeek · Model guide
DeepSeek V4.1 Flash
DeepSeek's Flash model in its V4.1 architecture series, now served as DeepSeek-V4.1-Flash-1017, an update released on 2026-10-17 with the same architecture and API model name as the September release. It is natively multimodal, with a 1M-token context window, up to 384K output tokens, and a thinking mode that can be turned off or dialled between low, high and max. DeepSeek says the 1017 update scores higher across its benchmarks while finishing tasks faster with shorter outputs, and it also serves the retired V4 Flash model names.
What this model is good at
What the vendor positions it for, and which of our models to reach for instead.
VendorDeepSeek calls 1017 a minor update to V4.1 Flash, the smallest model in its new architecture series, that improves reasoning, coding, agentic and visual benchmarks while cutting task run time and output length; it says the update was tuned in particular for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy.source
Terminal and coding agents · For the 1017 version the vendor reports Terminal-Bench 4.0 at 42.3 (Claude Code harness) and DeepSWE 1.1 at 77.0.source
Building whole projects from scratch · DeepSeek says self-feedback RL training improved the 1017 version at building complete projects such as games and 3D scenes from a short brief; it reports ProgramBench at 29.1 and NL2Repo-Bench at 68.2.(vendor claim)source
Agents that read screenshots and charts · Image understanding is native to this model rather than a separate experimental variant, and one request can carry up to 600 images.source
Hard knowledge and research · For the 1017 version the vendor reports Humanity's Last Exam at 42.6 without tools (44.0 on the text-only subset).source
High-throughput production traffic · DeepSeek documents a concurrency limit of 2500 for the flash tier, five times the V4 Pro limit.source
- deepseek/deepseek-v4-propredecessor —DeepSeek kept V4 Pro (DeepSeek-V4-Pro-0813) running as a separate model after 2026-09-14 instead of routing it to V4.1 Flash; it has no image input.source
- deepseek/deepseek-v4-flashpredecessor —V4 Flash and V4 Flash Vision Exp are retired upstream; those model ids are kept as compatibility aliases and are served by V4.1 Flash.source
Pricing and billing
Billed per token. Cached input is charged at the cache-read rate.
| Price / 1M tokens | You pay |
|---|---|
| Input | $ 0.30 |
| Output | $ 1.20 |
| Cache read | $ 0.0060 |
| Cache write | $ 0.30 |
+ 2.0K × $ 1.20
Capabilities and limits
What CrossModel guarantees across every route this model can take right now.
- Context window
- 1.0M tokens
- Max output
- 384.0K tokens
- Input / output modalities
- Text + Image → Text
- Streaming
- Supported
- Tool calling
- Supported
- Structured output (JSON)
- Supported
- Image input
- Supported
- Thinking mode
- Supported· can be turned off
- Reasoning effort
- low · medium · high · xhigh · max
- Available endpoints
- /v1/chat/completions · /v1/responses · /v1/messages
Published benchmarks
Scores the vendor reported, with the evaluation setup each one came from.
- Agents' Last Examscore·DeepSeek · 2026-10-17setup not fully disclosed
厂商自测（1017）；来源页未披露该项的评测设置。1017 来源：DeepSeek 更新日志 2026-10-17 条目，与 HuggingFace 模型卡 deepseek-ai/DeepSeek-V4.1-Flash-1017 的评测表一致。
35.4% - ApexBench v1.1score·DeepSeek · 2026-10-17setup not fully disclosed
厂商自测（1017）。来源页写作 Apex Bench v1.1，未说明 metric，保守记为 score。这是 DeepSeek 自 V4-Flash-Vision-Exp 起报告的 ApexBench（含多模态元素），不是 9 月首发版所录的 MathArena Apex；与 x-ai 的 APEX-Agents / APEX-SWE 是否同一测评未获证实，故单独立项。1017 来源：DeepSeek 更新日志 2026-10-17 条目，与 HuggingFace 模型卡 deepseek-ai/DeepSeek-V4.1-Flash-1017 的评测表一致。
72.1% - AutomationBench v1.0.6score·DeepSeek · 2026-10-17setup not fully disclosed
厂商自测（1017）。来源页写作 Automation-Bench v1.0.6，与目录既有的 AutomationBench v1.0.6 是同一身份（9 月首发版的 54.8 当时未标版本号，录在无版本的 AutomationBench 下）；未披露评测设置。1017 来源：DeepSeek 更新日志 2026-10-17 条目，与 HuggingFace 模型卡 deepseek-ai/DeepSeek-V4.1-Flash-1017 的评测表一致。
60.5%
- DeepSWE 1.1resolved_rate·DeepSeek · 2026-10-17Mini-SWE-V2setup not fully disclosed
厂商自测（1017）。来源页注明使用 Mini-SWE-V2 框架测试（harness 照抄来源页写法），未披露推理档位与采样设置。1017 来源：DeepSeek 更新日志 2026-10-17 条目，与 HuggingFace 模型卡 deepseek-ai/DeepSeek-V4.1-Flash-1017 的评测表一致。
77% - NL2Reposcore·DeepSeek · 2026-10-17setup not fully disclosed
厂商自测（1017）。来源页写作 NL2Repo-Bench，沿用目录既有的 NL2Repo 身份；只给出百分数，未说明具体 metric，保守记为 score；未披露评测设置。1017 来源：DeepSeek 更新日志 2026-10-17 条目，与 HuggingFace 模型卡 deepseek-ai/DeepSeek-V4.1-Flash-1017 的评测表一致。
68.2% - ProgramBenchscore·DeepSeek · 2026-10-17setup not fully disclosed
厂商自测（1017）；来源页只给出数值，未说明具体 metric，保守记为 score；未披露评测设置。1017 来源：DeepSeek 更新日志 2026-10-17 条目，与 HuggingFace 模型卡 deepseek-ai/DeepSeek-V4.1-Flash-1017 的评测表一致。
29.1%
- BabyVision (with tools)score·DeepSeek · 2026-10-17setup not fully disclosed
厂商自测（1017），来源页标注 w/tools，未披露具体工具集合；只给出数值，未说明具体 metric，保守记为 score。1017 来源：DeepSeek 更新日志 2026-10-17 条目，与 HuggingFace 模型卡 deepseek-ai/DeepSeek-V4.1-Flash-1017 的评测表一致。
92.5% - Chartography (with tools)score·DeepSeek · 2026-10-17setup not fully disclosed
厂商自测（1017），来源页标注 w/tools；与不带工具的 Chartography 不是同一身份，不要混用。来源页未披露具体工具集合。1017 来源：DeepSeek 更新日志 2026-10-17 条目，与 HuggingFace 模型卡 deepseek-ai/DeepSeek-V4.1-Flash-1017 的评测表一致。
83% - ZeroBench main (with tools)pass_at_5·DeepSeek · 2026-10-17setup not fully disclosed
厂商自测（1017）。来源页写作 ZeroBench-main (w/tools)，并注明使用 Pass@5 评测；未披露具体工具集合。与 V4-Flash-Vision-Exp 当年未标划分、未标工具的 ZeroBench 不是同一身份。1017 来源：DeepSeek 更新日志 2026-10-17 条目，与 HuggingFace 模型卡 deepseek-ai/DeepSeek-V4.1-Flash-1017 的评测表一致。
51%
Vendor-reported numbers, not CrossModel measurements. Scores are only comparable when the benchmark version, metric and evaluation setup match, so nothing here is averaged or ranked.
Use it in your tools
Point the base URL at CrossModel and paste this model ID — every tool below has a setup guide.
Related models
Other tiers from the same vendor, plus similarly priced alternatives.
Frequently asked questions
What is DeepSeek V4.1 Flash?
How much does DeepSeek V4.1 Flash cost?
Does DeepSeek V4.1 Flash support tool calling and structured output?
Which endpoint do I call?
Can I try it without writing code?
Chat first, integrate later
Chat first, then wire it in once you like the answers.