DeepSeek V4 Preview

DeepSeek V4 Preview is DeepSeek's latest open-source MoE language model with 1M token context support. On Atoms, you can select it as your AI model and build working apps by describing what you want in natural language. Atoms' multi-agent system uses V4 to handle everything from research to code generation to deployment.

V4-Pro has 1.6T total parameters (49B active) and delivers frontier-level reasoning, knowledge, and agentic performance. V4-Flash has 284B total parameters (13B active) with faster response times and lower API costs. Flash performs on par with Pro on simpler agent tasks, making it ideal for cost-sensitive or latency-critical applications.

DeepSeek V4 is available directly on atoms.dev. For reference, DeepSeek's own API prices V4-Pro at $1.74/M input tokens (currently 75% off until May 5, 2026) and V4-Flash at $0.14/M input tokens. Atoms integrates the model into its build workflow so you can focus on shipping rather than managing API billing.

Yes. DeepSeek V4 supports up to 1M tokens of context with a max output of 384K tokens. Its hybrid attention architecture reduces compute and memory costs dramatically — requiring only 27% of inference FLOPs and 10% of KV cache compared to V3.2 at the same context length. This makes processing entire repositories or long documents practical.

According to DeepSeek's published benchmarks, V4-Pro-Max scores competitively with Opus-4.6 Max and GPT-5.4 on coding (93.5 LiveCodeBench vs 88.8 for Opus), agentic tasks (80.6% SWE Verified), and math reasoning. It trails Gemini-3.1-Pro on some knowledge benchmarks but leads on several coding and shortlist evaluations.

Absolutely. Atoms supports full code export and GitHub sync, so everything you build with DeepSeek V4 is yours. The model itself is MIT-licensed and open-source with weights on Hugging Face. You can move your project to any hosting provider or continue developing independently at any time.