DeepSeek/V4.1 Flash
DeepSeek's Latest, Now Multimodal
A natively multimodal 552B-parameter model on a new, memory-efficient architecture that activates just 8B to 16B parameters per token. Put it to work in your browser, no setup required.
Model preset to DeepSeek V4.1 Flash · Sign up to unlock more models
Trending DeepSeek V4.1 Flash use cases
The hottest builds from the X community: one-shot Three.js worlds, games, simulations, and renders people are making with DeepSeek V4.1 Flash. Every clip below is a real post, with its prompt.
A one-shot 'Pagoda Universe' scene, spun up from a single prompt.
Build with DeepSeek V4.1 FlashA one-shot pagoda garden scene run locally on Blackwell GPUs at 200+ tokens per second, which the author called the best output they had gotten from a local model.
Build with DeepSeek V4.1 FlashSame prompt, two models: DeepSeek V4.1 Flash finished in 4 minutes for $0.14, versus Kimi K2.8 at 6 minutes and $0.91.
Build with DeepSeek V4.1 FlashA one-shot Three.js Mars rover sky-crane landing sequence, animated and looped, fully procedurally generated with no assets.
Build with DeepSeek V4.1 FlashOn the same task at highest reasoning, DeepSeek V4.1 Flash cost $0.17 against Kimi K3's $6.50.
Build with DeepSeek V4.1 FlashA one-shot Remotion video that recreates a Gemini 3 Pro ad almost perfectly, from just the reference video, with no images or follow-up prompts.
Build with DeepSeek V4.1 FlashA one-shot Three.js cutaway turbofan simulation that got the stators, rotors, and airflow right, from a single short paragraph prompt.
Build with DeepSeek V4.1 FlashA one-shot autonomous dungeon game whose pathfinding handled wall edges cleanly, something the author said usually takes other models two tries.
Build with DeepSeek V4.1 FlashThe viral one-shot pirate ship prompt, generated in about two minutes straight from the vanilla web UI with no agentic harness.
Build with DeepSeek V4.1 FlashA one-shot Three.js scene of an octopus band playing music for a crowd of fish, all in a single HTML file.
Build with DeepSeek V4.1 FlashA massive one-shot Three.js sunflower valley and village with mouse pan and zoom, built with high thinking over the API and running smoothly in the browser.
Build with DeepSeek V4.1 FlashA working SVG character rig that met the brief on the first try.
Build with DeepSeek V4.1 FlashA one-shot Three.js pastel sky world with a floating sky rail and a clock tower, single HTML, no external assets.
Build with DeepSeek V4.1 FlashA one-shot, procedurally generated infinite dune world in a single HTML file.
Build with DeepSeek V4.1 FlashA one-shot Three.js robotic arm arranging blocks to build a tower, single HTML file, no harness.
Build with DeepSeek V4.1 FlashA one-shot 3D chess board playing through a short premade game, single HTML file.
Build with DeepSeek V4.1 FlashWhat is DeepSeek V4.1 Flash?
DeepSeek V4.1 Flash is DeepSeek’s latest model, released on September 10, 2026. It is the smallest model in DeepSeek’s new architecture family and the first non-experimental DeepSeek with native visual understanding, so it reads images as well as text. DeepSeek positions it as its most capable model, and reports that tests by multiple parties put it ahead of the previous V4 Pro on performance, cost, speed, and total runtime; from September 14, 2026 the company routes V4 Pro requests to V4.1 Flash.
Under the hood it is a 552B-parameter mixture-of-experts model built on a new Causal Encoder-Decoder design that activates only about 8B parameters to read your input and 16B to write its output, cutting its KV-cache footprint sharply (about a quarter of the HBM and an eighth of the SSD of the previous generation). It is served through the DeepSeek API as “deepseek-flash”. On Happycapy, DeepSeek V4.1 Flash powers an agent inside a secure cloud sandbox that can write and run code, build sites, and deliver finished work from a single prompt, with no local setup.
How does DeepSeek V4.1 Flash perform on benchmarks?
DeepSeek V4.1 Flash compared with Kimi K3, GLM 5.3, Opus 5, and GPT-5.6 Sol on four agentic evals DeepSeek published at launch. Switch tabs to see each eval. All figures are DeepSeek-reported.
Competitor figures are drawn from the respective developers' published system cards or benchmark leaderboards.
Source: DeepSeek V4.1 Flash release
Full benchmark table
| DeepSeek V4.1 Flash | DeepSeek V4 Pro 0813 | DeepSeek V4 Flash 0731 | GLM 5.3 | Kimi K3 | GPT-5.6 Sol | Claude Opus 5 | |
|---|---|---|---|---|---|---|---|
| GPQA Diamond | 90.9 | 92.4 | 89.9 | 88.1 | 92.9 | 94.1 | 93.4 |
| HLE | 36.8 (39.1*) | 42.7* | 37.8* | 42.0* | 43.5 | 44.5 | 56.3 |
| Codeforces (Rating) | 3471 | 3348 | 3289 | — | — | — | — |
| MathArena Apex | 65.6 | 65.3 | 58.6 | — | 65.6 | — | — |
| Terminal-Bench 2.1 | 90.6 | 87.9 | 82.7 | 88.2 | 88.3 | 88.8 | 89.1 |
| Terminal-Bench 3.0 | 30.0 | 11.8 | 7.6 | 28.3 | 17.7 | 34.4 | 43.3 |
| Terminal-Bench 4.0 | 31.2 | 12.4 | 7.0 | 37.9 | 12.6 | 39.9 | 51.8 |
| DeepSWE v1.1 | 74.2 | 62.7 | 54.4 | 66.9 | 67.5 | 73.0 | 74.0 |
| ProgramBench | 20.3 | 15.5 | — | 19.0 | 17.5 | 23.0 | 37.0 |
| NL2Repo-Bench | 65.4 | 61.5 | 54.2 | 58.0 | 58.0 | 56.8 | 75.3 |
| CyberGym | 88.1 | 83.3 | 76.7 | 84.5 | 80.0 | 84.5 | — |
| SEC-Bench Pro | 62.8 | 56.4 | 30.9 | — | — | 74.3 | — |
| ExploitGym | 15.3 | 5.4 | 1.8 | 15.0 | — | 33.7 | 22.1 |
| HLE (w/ tools) | 63.9 | 60.0 | 51.5 | 62.5 | 59.8 | — | 63.6 |
| Automation-Bench | 54.8 | 43.2 | 37.7 | 48.8 | 46.7 | 45.8 | 50.3 |
| Agents' Last Exam | 31.8 | 25.7 | 25.2 | 28.5 | 27.6 | 26.7 | 28.6 |
| Chartography (w/ tools) | 78.9 | — | — | — | 68.1 | 79.9 | 84.0 |
| BabyVision (w/ tools) | 89.6 | — | — | — | 85.7 | 88.9 | 94.1 |
| ZeroBench-main (w/ tools) | 49.0 | — | — | — | 41.0 | 53.0 | 52.0 |
Source: DeepSeek V4.1 Flash release (deepseek.com). All figures are DeepSeek-reported; third-party scores are the best of self-reported or publicly available results. Best score per row in bold; a dash marks a result DeepSeek did not report. * denotes the text-only subset of HLE.
What DeepSeek V4.1 Flash is best at
Native multimodal understanding
The first non-experimental DeepSeek that can see: it reads images alongside text in one prompt, so it can work from screenshots, diagrams, and reference images, not just words.
Agentic coding
Writes, runs, and debugs real code end to end inside a sandbox, from a single component to a full build, driven by one prompt.
Memory-efficient long runs
Its Causal Encoder-Decoder design fires only a small slice of its 552B parameters per token and sharply cuts KV-cache memory, so long, tool-heavy agent runs stay affordable.
One-shot 3D and browser games
Community builders use it to spin up Three.js worlds, playable games, and simulations from a single prompt, often on the first try and for pennies.
Who is DeepSeek V4.1 Flash for?
Developers & vibe-coders
Anyone who wants to describe a build in plain language and have an agent write, run, and refine the code.
Multimodal builders
People working from screenshots, mockups, and reference images who need a model that can actually see the input.
Cost-sensitive, high-volume teams
Teams running many agent passes who want list-price flash-tier economics and memory-efficient inference.
DeepSeek users moving off V4 Pro
Anyone whose V4 Pro or V4 Flash workflow is being routed to V4.1 Flash and wants to run it agentically in the browser.
DeepSeek V4.1 Flash FAQ
What is DeepSeek V4.1 Flash?+
DeepSeek V4.1 Flash is DeepSeek's latest model, released on September 10, 2026. It is the smallest model in DeepSeek's new architecture family and the first non-experimental DeepSeek with native visual understanding, so it accepts both text and images. DeepSeek positions it as its most capable model.
Who makes DeepSeek V4.1 Flash?+
DeepSeek. It is the newest model in the DeepSeek line, and from September 14, 2026 DeepSeek routes its previous V4 Pro requests to V4.1 Flash.
Is DeepSeek V4.1 Flash multimodal?+
Yes. It is the first non-experimental DeepSeek with native visual understanding, so it can read images as well as text in a prompt, for example a screenshot, a diagram, or a reference image.
What is DeepSeek V4.1 Flash good at?+
It shines at agentic coding and one-shot 3D and browser-game builds: community testers use it to generate Three.js worlds, games, and simulations from a single prompt, often on the first try. Because it can also read images, it works from screenshots and reference shots as well as text.
What architecture does DeepSeek V4.1 Flash use?+
It is a 552B-parameter mixture-of-experts model built on a new Causal Encoder-Decoder design that activates only about 8B parameters to read input and 16B to generate output. DeepSeek reports it uses about a quarter of the HBM and an eighth of the SSD of the previous generation.
How much does DeepSeek V4.1 Flash cost?+
DeepSeek lists it at $0.30 per million input tokens and $1.20 per million output tokens on its official API, with off-peak pricing at half those rates. On Happycapy, DeepSeek V4.1 Flash is available on paid plans; you can create a free account to explore the platform first, then upgrade to run it.
How is DeepSeek V4.1 Flash different from V4 Pro?+
DeepSeek reports that tests by multiple parties put V4.1 Flash ahead of V4 Pro on performance, cost, speed, and total runtime, and from September 14, 2026 it routes V4 Pro requests to V4.1 Flash, so V4.1 Flash effectively supersedes the Pro tier.
How do I run DeepSeek V4.1 Flash on Happycapy?+
Type your task into the prompt box on this page (the model is already set to DeepSeek V4.1 Flash) and hit send. Happycapy spins up a secure cloud sandbox and lets the model build, code, or analyze the task end to end, then delivers the result. No installs or API keys required.
Build with DeepSeek V4.1 Flash now
Create a free account to explore Happycapy, then upgrade to put DeepSeek's latest multimodal model to work on real builds, right in your browser, no setup needed.
Get started for free