DeepSeek V4 Preview Released

[Updated July 21, 2026] This article covers the April 2026 preview. V4 has since reached general availability, and the legacy model IDs deepseek-chat / deepseek-reasoner retire on July 24, 2026, 15:59 UTC. See DeepSeek V4 Requirements Reference (2026 GA) for memory requirements, API pricing and migration steps.

Hangzhou-based DeepSeek released DeepSeek V4 Preview on April 24, 2026, publishing weights on Hugging Face. After their attention-grabbing model in 2025, V4 continues the open-weight Mixture-of-Experts (MoE) line. The same day, the API and chat.deepseek.com Expert / Instant modes went live.

Two MoE models shipped:

Both use MoE so only a subset of experts activates at inference time — runtime memory is far below the headline parameter count. Both support 1-million-token context, suiting long documents, large codebases, and multi-doc cross-reasoning.

V4-Pro and V4-Flash weights are posted to Hugging Face. As Preview releases, final license terms, commercial-use boundaries, and fine-tuning policy should be verified on the official repository. The strategy of offering near-frontier performance under open weights, contra OpenAI / Anthropic / Google, continues from the prior generation — making this a strong choice for teams pursuing local or self-hosted deployments.

Full-precision 1.6T MoE typically exceeds standard enterprise inference budgets; quantized weights and tuned serving stacks (vLLM, TGI) are the realistic baseline. V4-Flash (284B / 13B active) is more tractable — production-speed inference is feasible on A100 80GB or single H100 nodes. For consumer-grade hardware (e.g., Mac mini), V4 is impractical at full precision; pair quantized V4-Flash with smaller open models (Gemma 4, Qwen 3.6-27B).

We're building templates so OpenClaw can route lightweight tasks to V4-Flash in local / hybrid setups, while V4-Pro is reserved for cloud-API integration. See AI Consulting for tailored deployment advice.

V4 moved from preview to general availability in mid-July 2026. The lineup (V4-Pro 1.6T / 49B, V4-Flash 284B / 13B) and the 1M context are unchanged, and weights remain MIT-licensed. Three things changed in practice.

- Legacy IDs retire: deepseek-chat / deepseek-reasoner stop working after July 24, 2026, 15:59 UTC. Migrating to deepseek-v4-flash / deepseek-v4-pro is mandatory
- Reasoning is now a mode: Thinking / Non-Thinking replaces choosing a separate reasoning model
- Reported peak-hour pricing: 9:00–12:00 and 14:00–18:00 billed at twice the off-peak rate, per reporting (not yet reflected in the official pricing docs)

See DeepSeek V4 Requirements Reference (2026 GA) for memory requirements by quantization, GPU configurations, API rates and a migration checklist.

Q1: How does Preview differ from GA?
A: Preview ships before final licensing and benchmark validation. Specs and terms may shift. Production adopters may prefer to wait for GA.

Q2: Can V4-Pro run on a Mac mini?
A: Not at full precision, and even aggressive quantization is impractical at this size. On Mac mini-class hardware, prefer quantized V4-Flash or open models like Gemma 4 / Qwen 3.5-9B.

Q3: Is it usable commercially?
A: Verify the latest official license matches your requirements before commercial deployment.

- DeepSeek V4 Preview Release (DeepSeek API Docs, 2026/4/24)
- DeepSeek V4 — Simon Willison's blog (2026/4/24)
- China's DeepSeek drops new model (CNN, 2026/4/24)