MindStudio's public eight-task hands-on test gave V4-Pro-0813 a score of 61/80 (76.25%), showing strengths in frontend work, planning, mathematics, and long-horizon agents, while SVG work, game polish, and overthinking simple tasks remain boundaries.
Tasks it suits: Frontend generation, complex planning, long-horizon data-generation/fine-tuning/local-UI tasks, and agent work that requires clarifying ambiguities first.
Tasks it does not suit: One-line fixes, or situations where the model is expected to always answer briefly and avoid overhauling code; the article observed that Pro can overthink and over-engineer.
Applicable model version: DeepSeek-V4-Pro-0813 GA; the article explicitly compares it with the April preview.
Applicable client, agent, or API: The article does not disclose a standardized client; the results come from the author's coding/reasoning benchmark and hands-on use.
Recommended reasoning tier and parameters: Not disclosed; in an actual integration, the official low/high/max settings can be retested, but the effort setting cannot be inferred from the article's score.
Model/version: V4-Pro-0813; compared with V4 preview, V4 Flash, Kimi K3, Opus 5, Fable 5, Muse Spark 1.2, GLM 5.2, and others (using the names current in the article at the time).
Inputs/task types: Eight tasks covering a multi-elevator logic simulation, 3D interaction, a folding-table animation, an SVG panda, a bow-and-arrow game, permutation mathematics, long-horizon autonomous data generation/fine-tuning/local Web UI work, and a 3D dual-time-zone watch.
Evaluation method: Each task received 0–10 points. The article discloses the task types, per-task scores, and total score, but not the complete prompts, model-call parameters, or number of repetitions.
Temperature, thinking effort, maximum output, tool list, system prompt, random seed, and number of runs were not disclosed. The article also says that community specs list a 1M context, but this has not been confirmed by DeepSeek and cannot be treated as the configuration for this evaluation.
| Task | Score (/10) | Observation |
|---|---|---|
| Multi-elevator logic simulation | 6 | Core logic mostly works, but edge cases are incomplete |
| 3D contact lens case interaction | 8 | Comparable to top-tier models |
| Folding-table animation | 9 | Tied for the top tier with Fable 5, Kimi K3, GLM 5.2, and Sonnet 5 |
| SVG panda | 5 | A clear weakness, below Muse Spark 1.2's 10 |
| Bow-and-arrow game | 6 | Runnable but not polished enough |
| Permutation mathematics | 10 | Reached 2460, matching the highest score among multiple models |
| Long-horizon agent | 10 | Autonomously completed data generation, model fine-tuning, and a local Web UI with no human intervention |
| 3D dual-time-zone watch | 7 | The highest score in this test at the time, above V4 Flash's 6 and the previous Fable 5's 4 |
| Total | 61/80 (76.25%) | The article reports the preview at 24.8% |
Official/aggregate leaderboard scores (as reported in the article): Terminal Bench 2.1 87.9; Cyberjim 83.3; Automation Bench 31.8; HLE without tools 42.7, with tools 60. The article does not provide complete harnesses for these projects, so this note records them separately from the eight-task hands-on test.
The eight-task results support including V4-Pro-0813 among candidates for complex frontend and long-horizon agent work; its total score should not obscure its limitations in fine-grained SVG output, edge-case handling, and efficiency on simple tasks. The choice between Pro and Flash should be A/B tested according to “planning/complex tasks vs. everyday execution/speed,” rather than defaulting to Pro for everything.
Limitations: The article provides no original prompt for each task, tool trace, model configuration, number of repetitions, or blind evaluators; “independent” can only be understood as the author's hands-on test in a non-official article, not as a fully reproducible public dataset.
Reproduction steps: Recreate the eight task categories and publish the complete inputs; fix the model snapshot, effort, tools, context, and time limit; run each task at least three times; use the same 0–10 rubric for blind evaluation of functionality, visual quality, edge-case handling, and the amount of manual editing; run V4-Flash, Claude, Kimi, and others in the same harness.
Cost tracking: Record token costs under the official peak and off-peak prices at the same time, so quality advantages are not conflated with price advantages in a single score.
The MindStudio article gives per-task scores of 6/8/9/5/6/10/10/7, totaling 61/80, and reports 24.8% for the preview; the body also records firsthand observations of overthinking, over-engineering, frontend generation, planning, and ambiguity-clarification issues. This note does not treat the community rumor of a 1M context as a confirmed specification.
The sample contains only eight tasks, so task selection and subjectivity in scoring can significantly affect the total.
The HLE, Terminal, Cyberjim, and other figures in the article are not from the same experiment as the eight-task hands-on test and cannot be combined into an overall ranking.
The long-horizon task described as involving “no human intervention” still requires reviewing the run logs; the article does not disclose the complete trace.
The article's summary of the observed behavior is “overthinks simple problems” and “overengineering”; these two observations are more suitable than the total score for production acceptance tests.
DeepSeek V4 Pro
