DeepSeek V4.1-Flash
- Provider
- DeepSeek
- Status
- Current
- Context
- 1,000,000 tok
- Price
- $0.3 / $1.2 /MTok
DeepSeek V4.1-Flash is an open-weight model from DeepSeek, released on 10 September 2026 on the DeepSeek API as deepseek-flash and on Hugging Face under the MIT licence (DeepSeek). It replaced V4-Flash on the API, and DeepSeek says it is phasing out DeepSeek V4-Pro: from 04:00 UTC on 14 September 2026, all deepseek-v4-pro requests route to V4.1-Flash at V4.1-Flash prices until V4.1-Pro launches.
Artificial Analysis scores it 39 on the Intelligence Index v4.3.2 at max effort, level with GPT-5.5 and Grok 4.5, at $0.27 per completed index task. It costs $0.15 input and $0.60 output per million tokens off-peak, double that in peak hours. DeepSeek published no SWE-bench figure.
Quick specs
| Developer | DeepSeek |
| Released | 10 September 2026 |
| API name | deepseek-flash |
| Parameters | 552B total (Mixture-of-Experts); 8B active for input, 16B for output |
| Context window | 1,000,000 tokens |
| Max output | 384,000 tokens (DeepSeek API) |
| Input | Text and images |
| Price, off-peak | $0.15 input / $0.60 output per million tokens; cache hits $0.003 |
| Price, peak hours | $0.30 / $1.20; cache hits $0.006 |
| AA Intelligence Index | 39 (v4.3.2, max effort); 25 without reasoning |
| Cost per AA index task | $0.27 |
| Output speed | About 212 tokens a second (Artificial Analysis) |
| Licence | MIT, open weights |
What’s new
DeepSeek calls V4.1-Flash the smallest model in a new architecture family (DeepSeek).
A new architecture. The Hugging Face model card describes a Causal Encoder-Decoder design: a 40-layer Transformer split into a 20-layer causal encoder and a 20-layer decoder, with 552 billion parameters in total. It activates 8 billion parameters per token when reading input and 16 billion when generating output, against 13 billion for V4-Flash.
A smaller memory footprint. DeepSeek says the model’s KV cache, the memory a model holds for the text it has read, is a quarter the size of V4-Flash’s.
Native vision. V4.1-Flash reads images as well as text. V4-Flash had only an experimental vision variant, released on 21 August 2026 and retired with it.
Adjustable reasoning. The model card describes a reasoning effort that can be set anywhere from 1 to 100, trading computation for accuracy.
Benchmark performance
Independent: Artificial Analysis
| Measure | DeepSeek V4.1-Flash |
|---|---|
| Intelligence Index v4.3.2, max effort | 39 |
| Intelligence Index, no reasoning | 25 |
| Cost per index task | $0.27 |
| Output speed | About 212 tokens a second |
| Time to first token | 0.94 seconds |
| Output tokens to run the index | 250 million, against a median of 140 million |
Source: Artificial Analysis, measured on DeepSeek’s API. Its launch post scored the model 40 and said it overtakes DeepSeek V4-Pro 0813 as DeepSeek’s highest-scoring model while costing about four times less per token.
Vendor: DeepSeek’s launch table
DeepSeek’s figures, all at max reasoning effort, from the Hugging Face model card:
| Benchmark | V4.1-Flash | V4-Pro | V4-Flash | Claude Opus 5 | GPT-5.6 Sol | Kimi K3 | GLM-5.3 |
|---|---|---|---|---|---|---|---|
| Terminal-Bench 2.1 | 90.6 | 87.9 | 82.7 | 89.1 | 88.8 | 88.3 | 88.2 |
| DeepSWE v1.1 | 74.2 | 62.7 | 54.4 | 74.0 | 73.0 | 67.5 | 66.9 |
| GPQA Diamond | 90.9 | 92.4 | 89.9 | 93.4 | 94.1 | 92.9 | 88.1 |
DeepSeek also reports 36.8 on Humanity’s Last Exam, 54.8 on AutomationBench, 31.8 on Agent’s Last Exam and a Codeforces rating of 3,471. On DeepSeek’s table, V4.1-Flash leads every model listed on Terminal-Bench 2.1 and DeepSWE v1.1, and trails V4-Pro, Claude Opus 5, GPT-5.6 Sol and Kimi K3 on GPQA Diamond. These are vendor-run figures; the independent index places it well below those models.
Pricing
DeepSeek bills by time of day. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays; every other hour is off-peak at half the peak rate (DeepSeek).
| Per million tokens | Off-peak | Peak |
|---|---|---|
| Input, cache hit | $0.003 | $0.006 |
| Input, cache miss | $0.15 | $0.30 |
| Output | $0.60 | $1.20 |
V4-Flash cost $0.22 input and $0.66 output off-peak, so V4.1-Flash is cheaper on both. Requests to the retired deepseek-v4-flash and deepseek-v4-flash-vision-exp names, and since 14 September to deepseek-v4-pro, are served by V4.1-Flash and billed at these rates. The weights are free to self-host under the MIT licence.
How it compares
vs DeepSeek V4-Pro
DeepSeek says tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed and total runtime, and Artificial Analysis scored it above V4-Pro 0813. V4-Pro is a 1.6-trillion-parameter model to V4.1-Flash’s 552 billion. On DeepSeek’s own table, V4-Pro still scores higher on GPQA Diamond (92.4 against 90.9). See DeepSeek V4.
vs GPT-5.5 and Grok 4.5
GPT-5.5 and Grok 4.5 also score 39 on the Artificial Analysis index. V4.1-Flash reaches that score at $0.27 per completed task, against $2.63 for GPT-5.5 and $1.04 for Grok 4.5, and its weights are open.
vs Gemini 3.8 Flash
Gemini 3.8 Flash scores 41 on the index, two points higher, at $1.24 per completed task on introductory pricing of $0.75/$3.75 per million tokens that doubles on 1 January 2027. V4.1-Flash costs less per token and per task.
vs GPT-6 Luna
GPT-6 Luna, covered on the GPT-6 Sol page, scores 37 on the index at $0.10/$0.50 per million tokens and $0.07 per completed task. V4.1-Flash scores two points higher, and Luna costs less per token and per task.
Known limitations
No SWE-bench figure. DeepSeek published no SWE-bench Pro or Verified result, and its coding claims rest on Terminal-Bench 2.1 and DeepSWE v1.1, both vendor-run.
Verbose. Artificial Analysis measured 250 million output tokens to run its index against a median of 140 million, so cost per task rises faster than the token price suggests.
Peak-hour pricing. Rates double in peak hours, which fall in the European morning and the Asian working day.
Hosted-app and data caveats. DeepSeek’s hosted services store data in China and the app is banned or restricted in several countries; self-hosting the open weights avoids both. See the DeepSeek provider page.
Frequently asked questions
What is DeepSeek V4.1-Flash?
DeepSeek V4.1-Flash is an open-weight model from DeepSeek, released on 10 September 2026 under the MIT licence. It is a 552-billion-parameter Mixture-of-Experts model with a new encoder-decoder design, a 1-million-token context and image input, and it replaced V4-Flash on DeepSeek’s API as deepseek-flash.
How much does DeepSeek V4.1-Flash cost?
DeepSeek V4.1-Flash costs $0.15 per million input tokens and $0.60 per million output tokens off-peak, with cache hits at $0.003. In peak hours (01:00–04:00 and 06:00–10:00 UTC on weekdays) the rates double to $0.30, $1.20 and $0.006. The weights are free to self-host.
Is DeepSeek V4.1-Flash better than V4-Pro?
On the independent Artificial Analysis index, yes: it scored 40 at launch, above V4-Pro 0813, and DeepSeek says tests by multiple parties put it ahead of V4-Pro on performance, cost and speed. DeepSeek is phasing out V4-Pro, and since 14 September 2026 it serves deepseek-v4-pro requests with V4.1-Flash until V4.1-Pro launches.
Is DeepSeek V4.1-Flash open source?
The weights are published on Hugging Face under the MIT licence, so the model can be downloaded, self-hosted, fine-tuned and used commercially.
Can DeepSeek V4.1-Flash read images?
Yes. V4.1-Flash accepts text and image input and generates text. DeepSeek describes it as having native visual understanding.
Written on 30 September 2026, after the 10 September release. Artificial Analysis figures are independent; benchmark tables from DeepSeek are vendor-run.