Last updated:
DeepSeek V4.1 Flash: what it is, what it costs, how to use it
DeepSeek released V4.1 Flash on September 10, 2026 as the smallest model of its new architecture, then shipped an upgraded checkpoint, DeepSeek-V4.1-Flash-1017, on October 17. The API model name stays deepseek-flash and the price did not change. This page covers the specs, the peak and off-peak prices, DeepSeek’s benchmark claims and the quickest ways to use it.
DeepSeek V4.1 Flash at a glance
| Item | DeepSeek V4.1 Flash |
|---|---|
| Released | September 10, 2026; 1017 update on October 17, 2026 (DeepSeek-V4.1-Flash-1017) |
| API model name | deepseek-flash (old names deepseek-v4-flash and deepseek-v4-flash-vision-exp route to it at the same price) |
| Architecture | 552B mixture of experts, about 8B active parameters for input and 16B for output |
| Context / max output | 1M tokens / 384K tokens |
| Inputs | Text and images (native visual understanding) |
| Thinking | On by default at high effort; can be switched off per request |
| API formats | OpenAI-compatible and Anthropic-compatible, JSON output, tool calls, FIM (non-thinking only) |
| Concurrency limit | 2,500 |
| Weights and licence | deepseek-ai/DeepSeek-V4.1-Flash on Hugging Face, MIT licence |
Source: DeepSeek Models & Pricing
Source: DeepSeek V4.1 Flash release note
Source: Hugging Face model card
What changed in the 1017 update
On October 17 DeepSeek upgraded the model behind deepseek-flash, and the DeepSeek web and app to the same version. API calls and prices are unchanged; you get the new checkpoint without editing code.
- DeepSeek also says tasks finish faster and answers are shorter, without publishing figures for either.
- For agent work it recommends DeepSeek Harness 0.2.1-rc.1 or later.
- The weights on Hugging Face and the Artificial Analysis listing still describe the September release.
| Benchmark (DeepSeek’s numbers) | V4.1 Flash (Sept 10) | V4.1 Flash 1017 |
|---|---|---|
| Terminal-Bench 4.0 | 31.2 | 42.3 |
| DeepSWE v1.1 | 74.2 | 77.0 |
Source: DeepSeek API change log
Source: Hugging Face model card
DeepSeek V4.1 Flash price
Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday. Every other hour, weekends and Chinese public holidays are off-peak. In Beijing time the peak windows are 09:00–12:00 and 14:00–18:00.
| Per 1M tokens | Off-peak | Peak |
|---|---|---|
| Input, cache hit | $0.003 | $0.006 |
| Input, cache miss | $0.15 | $0.30 |
| Output | $0.60 | $1.20 |
Web Search and Fetch tools cost $0.75 per 1,000 successful requests each; through DeepSeek Harness, Claude Code or Codex they are free until November 1, 2026 (Beijing time).
Full API pricing, peak-hour table and code
Source: DeepSeek Models & Pricing
Benchmarks (DeepSeek’s table)
DeepSeek’s model card compares the September release with its older models and two Western flagships, all at maximum reasoning effort. Treat these as the vendor’s own results.
| Benchmark | DeepSeek V4.1 Flash | DeepSeek V4 Pro | DeepSeek V4 Flash | Claude Opus 5.0 | GPT-5.6 Sol |
|---|---|---|---|---|---|
| GPQA Diamond | 90.9 | 92.4 | 89.9 | 93.4 | 94.1 |
| Humanity's Last Exam (no tools) | 36.8 | 42.7 † | 37.8 † | 56.3 | 44.5 |
| Humanity's Last Exam (with tools) | 63.9 | 60.0 | 51.5 | 63.6 | — |
| Codeforces (rating) | 3471 | 3348 | 3289 | — | — |
| Terminal-Bench 2.1 | 90.6 | 87.9 | 82.7 | 89.1 | 88.8 |
| Terminal-Bench 4.0 | 31.2 | 12.4 | 7.0 | 51.8 | 39.9 |
| DeepSWE v1.1 | 74.2 | 62.7 | 54.4 | 74.0 | 73.0 |
| NL2Repo-Bench | 64.0 | 61.5 | 54.2 | 75.3 | 56.8 |
| AutomationBench | 54.8 | 43.2 | 37.7 | 50.3 | 45.8 |
| Agent's Last Exam | 31.8 | 25.7 | 25.2 | 28.6 | 26.7 |
| BabyVision (with tools) | 89.6 | — | — | 94.1 | 88.9 |
† Text-only subset of Humanity’s Last Exam. — means DeepSeek listed no result. The 1017 update scores higher on Terminal-Bench 4.0 and DeepSWE (see above).
Source: Hugging Face model card
Independent tests
Artificial Analysis lists V4.1 Flash (September release, Max effort) at 39 on its Intelligence Index, with 175.9 output tokens per second, 1.23 s to first token and about $0.27 per index task at the peak price. It used 250M output tokens to run the index, so it is a heavy thinker: off-peak hours halve that bill.
DeepSeek V4.1 Flash vs Gemini 3.8 Flash vs Claude Haiku 5.5
Source: Artificial Analysis: DeepSeek V4.1 Flash
How to use DeepSeek V4.1 Flash
- In a browser or app. The DeepSeek web chat and mobile app run the 1017 version since October 17.
- API (OpenAI format). Point any OpenAI SDK at https://api.deepseek.com and use model deepseek-flash.
- API (Anthropic format). Use https://api.deepseek.com/anthropic as the base URL with the Anthropic SDK.
- Claude Code. Set a few environment variables and Claude Code runs on DeepSeek; DeepSeek publishes the exact list.
- Self-host. Download deepseek-ai/DeepSeek-V4.1-Flash from Hugging Face. At 552B parameters it needs a multi-GPU server, not a desktop card.
API setup, Python code and Claude Code variables
Source: DeepSeek API change log
Source: DeepSeek API docs: first call
Source: DeepSeek Anthropic API guide
Source: DeepSeek guide: Claude Code
Source: Hugging Face model card
Is it open source? Can you use it commercially?
The repository and weights are released under the MIT licence, which allows commercial use, modification and redistribution with the licence notice kept. API use is covered by DeepSeek’s platform terms instead.
Source: Hugging Face model card
FAQ
What is DeepSeek V4.1 Flash?
DeepSeek’s smallest model in its new architecture, released September 10, 2026: a 552B mixture-of-experts model with a 1M context, image input and MIT-licensed weights. The API name is deepseek-flash.
What is DeepSeek V4.1 Flash 1017?
The October 17, 2026 update (DeepSeek-V4.1-Flash-1017). It replaced the model behind deepseek-flash at the same price; DeepSeek reports Terminal-Bench 4.0 at 42.3 and DeepSWE v1.1 at 77.0.
How much does DeepSeek V4.1 Flash cost?
$0.15 / $0.60 per 1M input/output tokens off-peak and $0.30 / $1.20 at peak; cached input is $0.003 off-peak and $0.006 at peak.
When are DeepSeek peak hours?
01:00–04:00 and 06:00–10:00 UTC, Monday to Friday (09:00–12:00 and 14:00–18:00 Beijing time). Weekends and Chinese public holidays are off-peak.
Is DeepSeek V4.1 Flash open source?
Yes, the weights are on Hugging Face under the MIT licence, which permits commercial use.
What is the context window?
1M tokens of context and up to 384K output tokens per request.
Independent guide, not affiliated with DeepSeek, Google or Anthropic. Prices and model versions change often; check the linked official pages before you build on them.