DeepSeek V4 Pro
A flagship coding model released by DeepSeek, recommended for use as a backend provider with Claude Code. It is noted for offering strong price-to-performance characteristics among coding-oriented large language models.
Core Facts
DeepSeek V4 Pro adopts a MoE (Mixture of Experts) architecture, activating only a subset of parameters during inference to reduce computational costs.
DeepSeek is deploying native multimodal capabilities at the Flash level rather than the flagship version first, with the intent to prioritize capturing high-frequency, large-scale application scenarios.
Interfaces that previously called DeepSeek V4 Pro will automatically switch to the new V4.1 Flash model, requiring no code changes from developers.
Timeline (last 90 days)
Gemini 4 was closer to a runnable state on its first generation, whereas DeepSeek V4 Pro's code, though longer, made elementary mistakes in basic logic.
In a non-strictly controlled test, the uploader asked Gemini 4 and DeepSeek V4 Pro to write a webpage in native JS that simulates the physics calculations (not animation) of a Tesla flying car scenario, with two gameplay modes: ramp flip and jumping over five cars, with each getting only one initial generation attempt.
This test is a non-strictly controlled experiment, and prompts, environment and API calling methods all affect the results; the uploader put the prompts in the comments section to encourage others to reproduce it.
In the ramp flip task, DeepSeek V4 Pro took about six and a half minutes to generate code, which was longer and had a complete interface, but after clicking start the car showed as crashed without moving; the cause was that the initial position height was exactly equal to the ground, and the code's rule that "touching the ground counts as landing" caused a false landing on the first frame.
In the jump-over-five-cars task, DeepSeek V4 Pro rewrote the entire code after the problem was pointed out, but it still crashed immediately after clicking start; its small model B4.1 Flash could climb and flip in the ramp flip but crashed at the end, and could only crawl slowly when jumping over five cars.
In the first task, a 3D Subway Surfers clone, the version generated by DeepSeek 4.1 Flash ran smoothly, with jumping, left-right movement, sliding and other operations free of bugs, and building modeling detail on par with the previous generation V4 Pro.
With DeepSeek V4 Pro as the backbone, SwarmRSI achieves a medal rate of 81.8% on MLEBench Lite.
With DeepSeek V4 Pro as the backbone, SwarmRSI's proportion matching SOTA on NatureBench is 21.1 percentage points higher than Claude Code with the same backbone.
With DeepSeek V4 Pro as the backbone, SwarmRSI achieves an average score of 74.2 on OpenProblems, exceeding the previous SOTA of 67.1.
For chat, summarization and Agent subtask suggestions, the DeepSeek Flash model is recommended; for code and reasoning tasks, the Pro model is recommended.
40 more timeline events
All Facts (20)
DeepSeek V4 Pro adopts a MoE (Mixture of Experts) architecture, activating only a subset of parameters during inference to reduce computational costs.
80%VerifiedDeepSeek is deploying native multimodal capabilities at the Flash level rather than the flagship version first, with the intent to prioritize capturing high-frequency, large-scale application scenarios.
75%VerifiedInterfaces that previously called DeepSeek V4 Pro will automatically switch to the new V4.1 Flash model, requiring no code changes from developers.
65%VerifiedFor advanced obfuscation techniques such as control flow flattening, AI's analytical capabilities may be significantly diminished.
65%VerifiedDeepSeek V4 Pro is a new model released in late April 2025.
65%VerifiedDeepSeek officially launched DeepSeek Harness (abbreviated as DSH) following the release of the DeepSeek V4 Pro model.
70%UnverifiedV4 Pro的代码生成速度明显慢于V4 Flash
90%UnverifiedDeepSeek officially announced that V4 Pro scored as high as 87.9 on the Terminal Bench benchmark.
60%UnverifiedCurrent AI-assisted game development is primarily centered on Web-based lightweight approaches, using frameworks such as HTML5 Canvas, Phaser.js, and Three.js.
60%UnverifiedGemini 4 was closer to a runnable state on its first generation, whereas DeepSeek V4 Pro's code, though longer, made elementary mistakes in basic logic.
50%UnverifiedThis test is a non-strictly controlled experiment, and prompts, environment and API calling methods all affect the results; the uploader put the prompts in the comments section to encourage others to reproduce it.
50%UnverifiedIn the first task, a 3D Subway Surfers clone, the version generated by DeepSeek 4.1 Flash ran smoothly, with jumping, left-right movement, sliding and other operations free of bugs, and building modeling detail on par with the previous generation V4 Pro.
50%UnverifiedWith DeepSeek V4 Pro as the backbone, SwarmRSI achieves an average score of 74.2 on OpenProblems, exceeding the previous SOTA of 67.1.
50%UnverifiedWith DeepSeek V4 Pro as the backbone, SwarmRSI achieves a medal rate of 81.8% on MLEBench Lite.
50%UnverifiedWith DeepSeek V4 Pro as the backbone, SwarmRSI's proportion matching SOTA on NatureBench is 21.1 percentage points higher than Claude Code with the same backbone.
50%UnverifiedFor chat, summarization and Agent subtask suggestions, the DeepSeek Flash model is recommended; for code and reasoning tasks, the Pro model is recommended.
50%UnverifiedThe three models involved in this comparison are DeepSeek V4 Pro, Claude Fable 5.1, and Grok Build 0.1.
50%UnverifiedThe free interfaces of MiniMax M2.7 and DeepSeek V4 Pro on the NVIDIA platform have been deprecated, leaving only the partner's paid interfaces.
50%UnverifiedOfficial model card scores for Qwen3.8-27B: GPQA Diamond 89.2, LiveCodeBench v6 90.3, SWE-bench Pro 61.7, Terminal Bench 2.1 73.0; Artificial Analysis Intelligence Index score of 52, on par with GPT-5.6 Luna, and just one point below GLM-5.2 (753B) and DeepSeek V4 Pro (1.7T).
50%UnverifiedOn the DeepSeek V4 Pro workload, with an interaction target of 160 tokens/second per user, the per-megawatt AI factory throughput of Vera Rubin-NVL72 is up to 30 times higher than that of GB300-NVL72.
50%