DeepSeek officially released DeepSeek V4 Preview on July 24, with both models open-sourced and a standard 1 million token context window. On the SWE-bench coding benchmark, V4 scored 83.7%, surpassing Claude Opus 4.6 (80.9%) and GPT-5.2 (80%), though it hasn't yet caught up with the latest flagship models.
V4-Pro: 1.6 Trillion Parameters, MoE Architecture
The V4 series includes two models: V4-Pro (1.6 trillion parameters, activates 49 billion per inference) and V4-Flash (284 billion parameters, activates 13 billion). Both use a Mixture-of-Experts (MoE) design that selectively activates only a subset of "expert" neurons per query, drastically cutting compute costs without sacrificing output quality. The 1 million token context can digest the entire Harry Potter series plus a large codebase in one go.
Coding Leadership, Agentic Performance Close to Top Tier
Beyond coding, DeepSeek-V4-Pro Max approaches GPT-5.4, Gemini-3.1-Pro, and Opus-4.6 in knowledge reasoning, and excels in agentic tasks. On SWE Verified it scores 80.6, nearly matching Gemini-3.1-Pro and slightly trailing Opus-4.6; on BrowseComp it achieves 83.4, just behind Gemini's 85.9; and on MCPAtlas Public it hits 73.6, close to Opus-4.6's 73.8 and ahead of GPT-5.4 and Gemini. However, on several difficult reasoning leaderboards, V4 still lags behind GPT-5.4, Gemini-3.1-Pro, or Opus-4.6.
Cost Dominance: Input Price 27x Cheaper Than Closed Rivals
DeepSeek V4 charges approximately $0.3 per million input tokens — about 27 times cheaper than comparable closed-source models. V4-Pro's pricing may drop further in the second half of the year as Huawei Ascend 950 chips become widely available.
Fully Runs on Huawei Ascend 950
V4 operates entirely on Huawei's Ascend 950 PR chips, bypassing Nvidia's CUDA ecosystem. Amid tightening US export controls, this is both an engineering decision and a long-term survival strategy. DeepSeek has delivered the first production-grade large model that runs exclusively on domestic Chinese chips.