NEW · V4.1-Flash-1017 (released 10-17)
DeepSeek-V4.1-Flash-1017 usesRLSF self-feedback reinforcement learningfor training, featuring "Wish-Driven Programming"—a single wish-style description directly produces a complete project. DeepSWE v1.1 Pass@177.0%, with an average cost per task of only about$0.20(low tier $0.10), a major improvement over the previous generation's 74.2% / $0.50. API call namedeepseek-flash: 1M context, 384K max output, native vision input, with thinking mode enabled by default.
V4.1-Flash (09-10)
MoE with 552B total parameters and native multimodality (image input), approaching or even surpassing V4-Pro (0813) on multiple benchmarks; output price reduced by about 70% versus V4-Flash, with cache-hit input as low as $0.003/million tokens.
V4 series (04-24)
V4-Pro (1.6T total parameters / 49B activated) and V4-Flash (284B / 13B) are released as open weights under the MIT license, with 1M context + 384K output, and long-context compute overhead at only 27% FLOPs / 10% KV of V3.2.
The DeepSeek V4 family brings together frontier capabilities and extremely low inference costs: V4.1-Flash-1017 uses RLSF reinforcement learning to push agentic coding costs down to about $0.20/task; V4-Pro scores 80.6% on SWE-Bench Verified and 93.5% on LiveCodeBench; all weights are open under the MIT license, with a 1M token context and 384K max output. Try it for free below — or check out the benchmark data first.
No sign-up, no account required. Enter a prompt and get streaming responses.
The response will stream here…
This playground is powered by the DeepSeek large language model. Enter a prompt to get real-time streaming responses.
Four capabilities define the current DeepSeek generation — officially reported scores, reproducible from MIT open weights.
1M token context + 384K max output. Hybrid attention lets 1M workloads consume only about 27% of V3.2's FLOPs and 10% of the KV cache — long context no longer means a high bill.
V4.1-Flash-1017 trains project development capabilities end-to-end with Reinforcement Learning from Self-Feedback (RLSF): the model uses its own confidence and self-critique as reward signals, reaching 77.0% on DeepSWE v1.1.
The V4.1-Flash series natively accepts image input with no external vision module required — screenshots, UIs, and document images go straight into the same context and tool-calling loop.
V4-Pro and V4-Flash weights are openly available on Hugging Face under the MIT license (Flash is about 160GB), self-hostable, cleared for commercial use, with no regional restrictions, and the API is compatible with the OpenAI / Anthropic SDK.
Data sourced from DeepSeek official releases and public evaluations (as of October 2026).
| Evaluations | Models | Results | Description |
|---|---|---|---|
| DeepSWE v1.1 (Pass@1) | V4.1-Flash-1017 | 77.0% | Average cost per task is approximately $0.20, with the low tier at $0.10 |
| SWE-Bench Verified | V4-Pro | 80.6% | Fixing real-world repository-level issues |
| LiveCodeBench | V4-Pro | 93.5% | Real-time competitive programming |
| Terminal-Bench 2.0 | V4-Pro | 67.9% | Terminal agent tasks |
| Codeforces (Rating) | V4-Pro | 3,206 | Competitive programming Elo |
| Putnam-200 (Pass@8) | V4-Flash | 81.0% | Math competitions |
For the 16 cross-comparisons with models such as GLM-5.3, GPT-5.6, and Claude, seeBenchmarks / DeepSeek vs GLM。
Two-tier MoE + Compressed Sparse Attention keeps the compute bill for a 1M context as low as possible.
V4-Pro has 1.6T total parameters with 49B activated per token; V4-Flash has 284B total parameters with 13B activated; V4.1-Flash at 552B features an all-new encoder design. Tiers—and their costs—switch through the same API.
A combination of Compressed Sparse Attention (CSA) + Heavily Compressed Attention (HCA): at 1M context, FLOPs are only about 27% of V3.2's and the KV cache about 10%, optimizing long-context retrieval quality and cost at the same time.
Pre-trained on 32T+ tokens with FP4+FP8 mixed precision; V4.1-Flash-1017 adds RLSF post-training: using the model's own confidence/self-critique as the reward signal, with no reliance on human evaluators.
Entire codebases, contracts hundreds of pages long, and complete agent trajectories all fit in a single load; the 384K max output is enough to generate full projects and long reports. Thinking mode is enabled by default.
deepseek-flash has a concurrency limit of 2500, deepseek-v4-pro 500; the API is compatible with OpenAI and Anthropic SDKs, and supports JSON output and tool calling. For details, see1M context explained。
The main models currently sold via API — see the full catalog atAll models。
New · Recommended
552B MoE · native vision · RLSF · 1M/384K
Call name deepseek-flash
$0.15 / $0.60 per 1M tokens (off-peak)
1.6T/49B flagship · MIT · 1M/384K
API name deepseek-v4-pro
$0.66 / $1.98 per 1M tokens (off-peak)
552B · Native vision · Released 09-10
Output price about 70% lower than V4-Flash
Cache-hit input $0.003
284B/13B · MIT · ~160GB of weights
Putnam-200 81.0% · SWE-Bench 79.0%
$0.14 / $0.28 per 1M tokens
"Wish-Driven Programming": complete projects from a single-sentence description—official demos include a 3D model of the Forbidden City, a web-based racing game, a Godot recreation of Portal 2, and the 3D scene "Valley of Ages".
A 1M context window holds an entire code repository or hundreds of pages of documents; cache hits cost $0.003 per million tokens, driving the cost of RAG pipelines that repeatedly retrieve the same corpus down to nearly zero.
MIT open weights + OpenAI/Anthropic-compatible API: get frontier coding and reasoning capabilities without your data ever leaving the internal network, and migrating costs just one line of base_url.
For heavy reasoning and complex multi-file coding, choose V4-Pro; for high-concurrency, long-document, or cost-sensitive scenarios, choose deepseek-flash (now pointing to V4.1-Flash-1017, with vision support and 2500 concurrency).
The API is pay-as-you-go with no free quota; the weights are MIT open-source, so you can deploy them yourself. The playground on this site offers free streaming demos.
Calls to v4-pro are routed to V4.1-Flash starting 2026-09-14; the official recommendation is to use deepseek-flash directly for new projects. For details, seeFAQ。
Head back to the playground at the top of the page and enter a prompt to get a streaming response.