DeepSeek V4

DeepSeek released two new models as part of V4: Pro and Flash.

Both models feature a cost-effective 1 million context length.

DeepSeek V4 Pro has 1.6 trillion total parameters and 49 billion active parameters.

DeepSeek V4 Flash is a faster, cheaper model with 284 billion total parameters and 13 billion active parameters, offering near-pro reasoning on simpler agent tasks.

DeepSeek claims V4 Pro is the top open-source model in reasoning, STEM, coding, forensic workflows, and world knowledge, rivaling leading closed-source models.

DeepSeek claims V4 is on the level of Opus 4.5 on real-world agentic coding tasks and in certain cases beats Opus 4.6.

The author believes DeepSeek V4 is benchmark-maxed, meaning the team optimized heavily for benchmarks in their favor.

In benchmarks, DeepSeek claims V4 beats Claude Opus 4.6 Max, Gemini 3.1 Pro, and GPT 5.4 High in certain areas, but the author does not agree.

In the author's testing, Kim K 2.6, Quen, and Minia Max N2.7 outperform DeepSeek V4 preview in extended thinking and real-world use cases.

The current V4 is a preview release and an official version could fix current drawbacks.

The model has strong long context performance, extremely low cost, and is open-source under the MIT license.

From the author's testing, the model feels subpar, sloppy, and lazy in execution.

In a browser-based Mac OS clone test, DeepSeek V4 generated a lackluster, basic output without SVG icons or proper UI thinking, failing to capture macOS structure.

DeepSeek V4 Flash sometimes performs better than Pro in certain prompting cases but still delivers lackluster results.

DeepSeek V4 Pro failed to finish generating an off-road EV durability test, while Miniax M2.7 completed it with interactive car movement.

In a SAS landing page generation, DeepSeek V4 Pro's quality resembled GPT-3.5 output.

In a Slack clone generation, DeepSeek V4's output mimicked some structure but did not look like the actual app, while GLM 5.1 captured the Slack tone and style accurately.

In SVG generation, DeepSeek V4 Pro produced a decent butterfly but Quen 3.6 Plus outperformed it in almost every fashion.

For a 3D PS5 controller task, DeepSeek V4 Pro generated an object resembling a table, while Quen produced a recognizable controller.

In an Instagram feed clone, MiniAX M2.7 replicated the feed correctly, while DeepSeek V4 Pro was buggy and failed to compile.

On Code Arena, DeepSeek V4 ranks number three behind GLM 5.1 and Kim K 2.6, not number one.

DeepSeek V4 Pro pricing is 14 dollars per 1 million input tokens and 348 dollars per 1 million output tokens.

DeepSeek V4 Flash pricing is 3 cents per 1 million input tokens and 28 cents per 1 million output tokens.

The author states DeepSeek V4 is nowhere near other Chinese models like Miniax or Kim K 2.6.

Just because the model is cost-efficient does not automatically make it great; cheaper does not mean better.