V4.1 Flash
DeepSeek · released Sep 10, 2026
DeepSeek V4.1 Flash is a sparse mixture-of-experts model from DeepSeek, and the first built on the company's Causal Encoder-Decoder (CED) architecture. It activates 8B parameters on input and 16B on...
Specification
- Context window
- 1.05M
- Max output
- 944K
- Knowledge cutoff
- Not stated
- Parameters
- 763.2B
- Licence
- mit
- Serving providers
- 30
- Moderated
- No
- Uptime
- 100.0%
Intelligence
39.5
90th percentile
Coding
—
Coding Index
Agentic
—
Agentic Index
Output speed
296 t/s
Median across providers
Latency
306ms
Time to first token
Cost per task
$0.07
Estimated
Benchmarks
Where the score comes from
The Intelligence Index is a composite. These are the underlying evaluations this model was actually measured on.
Evaluation scores
Percentage correct · higher is better
- AA-LCR (long context)84.0%
- SciCode51.9%
- Humanity's Last Exam39.2%
An evaluation missing from this list was not run for this model — it is not a zero.
View as table
| Evaluation | Score |
|---|---|
| AA-LCR (long context) | 84.0% |
| SciCode | 51.9% |
| Humanity's Last Exam | 39.2% |
Against its peers
Intelligence Index · this model highlighted, nearest peers in grey
- Claude Opus 4.7Anthropic40.7
- 39.9
- Muse Spark 1.2Meta39.6
- V4.1 FlashDeepSeek39.5
- Gemini 3.7 FlashGoogle39.1
- GPT-5.4OpenAI39.0
- Grok 4.5xAI38.8
- GPT-5.5OpenAI38.4
Peers are the models sitting closest on the Intelligence Index — the set you would realistically choose between.
View as table
| Model | Intelligence |
|---|---|
| Claude Opus 4.7 | 40.7 |
| Qwen3.8 2.4T A95B | 39.9 |
| Muse Spark 1.2 | 39.6 |
| V4.1 Flash | 39.5 |
| Gemini 3.7 Flash | 39.1 |
| GPT-5.4 | 39.0 |
| Grok 4.5 | 38.8 |
| GPT-5.5 | 38.4 |
Percentile among all indexed models
Pricing
What it costs to run
List prices per million tokens, plus what one representative task works out to.
List price
- Input / 1M tokens
- $0.014
- Output / 1M tokens
- $0.891
- Cached input / 1M
- $0.005
- Blended 3:1
- $0.233
One task, estimated
$0.07
- Input tokens
- 50,000
- Output tokens
- 80,000
- Profile
- Reasoning
Estimated from list pricing: 50K input tokens plus 80K output tokens for reasoning models (25K for non-reasoning).
Serving providers
- OpenInference
- Relace
- Morph
- Decart
- Sail Research
- InferenceNet
- Cloudflare
- Wafer
- DeepInfra
- StreamLake
- CoreWeave
- GMICloud
- AtlasCloud
- Novita
- Phala
- DigitalOcean
- Ionstream
- Makora
- Crusoe
- PrimeIntellect
- Baidu
- BaseTen
- Alibaba
- Together
- SiliconFlow
- Modal
- Parasail
- Fireworks
- Venice
- DeepSeek
Speed and latency figures are medians across these providers, so a widely-served model reports a blend rather than any single endpoint.
Arena
Head-to-head generation quality
Elo from pairwise judgements, broken out by the kind of thing the model was asked to build.
- Overall Elo
- 1,357
- Win rate
- 58.2%
- Strongest at
- 3D scenes
- Tournaments
- 5,283
Elo by category
Dot position on a 1,225–1,375 scale · Elo has no meaningful zero
- 1,357
- 1,333
- 1,318
- 1,311
- 1,293
- 1,243
Agent categories (full-stack apps, mobile apps) are only scored for models tested in agent harnesses.
View as table
| Category | Elo | Win rate |
|---|---|---|
| 3D scenes | 1357 | 58.2% |
| Game development | 1333 | 54.5% |
| UI components | 1318 | 52.8% |
| Websites | 1311 | 50.7% |
| Data visualisation | 1293 | 50.4% |
| SVG | 1243 | 49.0% |