DeepSeek V4 - Nearly Frontier-Level, at a Much Lower Price

DeepSeek has released preview models V4-Pro and V4-Flash. Both are MoE models supporting a 1 million token context, distributed under the MIT license.

• V4-Pro: 1.6T total parameters, 49B active
• V4-Flash: 284B total parameters, 13B active
• Hugging Face sizes: Pro 865GB, Flash 160GB
• V4-Pro is larger in scale than Kimi K2.6, GLM-5.1, and DeepSeek V3.2, appearing to be the new largest open-weight model.

Pricing is central to this release. Flash is priced at $0.14 per 1 million input tokens and $0.28 for output, while Pro is $1.74 for input and $3.48 for output. Flash ranks among the cheapest in the small model class, and Pro among the cheapest in the large frontier model class.

Long-context efficiency has also improved significantly. At 1 million tokens, Pro's per-token FLOPs drop to 27% and KV cache to 10% compared to V3.2, while Flash drops to 10% FLOPs and 7% KV cache.

In its own benchmarks, Pro is competitive with frontier models, though slightly below GPT-5.4 and Gemini-3.1-Pro. The paper presents V4-Pro-Max, a reasoning-token-extended version, as surpassing GPT-5.2 and Gemini-3.0-Pro on some standard reasoning benchmarks.

In summary, this is an open-weight update that boosts both price and efficiency simultaneously.