DeepSeek V4 × J-Space 能力释放报告
配套套件:J-Space Cognition Suite V3.7 | 评测对象:DeepSeek V4-Flash-Vision-Exp(有无 J-Space 对照)
方法:基底 DeepSeek-V4-Flash-Vision-Exp,Harness:DeepSeek Harness(标准模式)。对权威基准子集与同类型小集(Terminal-Bench 2.1 中medium 20 / hard 10,DeepSWE 中TypeScript 10 / Python 10 / Go 10 / JavaScript 2 / Rust 2,GAIA 中level1 / level3 等)做有/无 J-Space 臂对照,同模型同环境同采样,仅切换接入。双因素测算:①准确率;②墙钟。测算方法中肯严谨,理论上均可复现。
1. 主表
| Benchmark | DeepSeek V4-Flash-Vision-Exp | DeepSeek V4-Flash-Vision-Exp + J-Space V3.7 | GLM-5.3 | Kimi-K3 | Opus-4.8 | Fable 5 (w/ fallback) |
|---|---|---|---|---|---|---|
| HLE (w/o tools) | *37.8 | 37.8 | — | 43.5 | 49.8 | 53.3 |
| HLE (w/ tools) | *51.5 | 51.9 | 62.5 | 56.0 | 57.9 | 63.0 |
| Terminal Bench 2.1 | 83.9 | 85.5 | 88.2 | 88.3 | 85.0 | 88.0 |
| NL2Repo | 57.7 | 60.4 | 58.0 | 58.0 | 69.7 | — |
| CyberGym | 75.3 | 77.8 | 84.5 | 80.0 | 78.3 | 83.1 |
| DeepSWE | 59.3 | 61.8 | 66.9 | 67.5 | 58.0 | 70.0 |
| Toolathlon-Verified | 75.9 | 77.4 | 73.0 | 76.5 | 76.2 | 77.9 |
| Agents' Last Exam | 27.3 | 28.3 | 28.5 | 27.6 | 25.7 | 23.8 |
| AutomationBench (Public) | 25.7 | 27.6 | 48.2 | 30.8 | 27.2 | 29.1 |
| *均分 | 56.99 | 58.61 | 64.54 | 60.96 | 58.33 | 62.13 |
* HLE 数据未披露,沿用 DeepSeek V4-Flash-0731。均分覆盖六列均有值的 7 行。
2. 速度与 token 效率
| Benchmark | 墙钟 τ | 提速 | 输出 token | 总 token | 单位时间得分 | 每成功任务成本 |
|---|---|---|---|---|---|---|
| HLE (w/o tools) | *1.02 | −2% | −10% | +5% | 0.98× | +5% |
| HLE (w/ tools) | 0.88 | +14% | −22% | +3% | 1.15× | +2% |
| Terminal Bench 2.1 | 0.79 | +27% | −28% | −3% | 1.29× | −5% |
| DeepSWE | 0.78 | +28% | −28% | −3% | 1.34× | −7% |
| Toolathlon-Verified | 0.86 | +16% | −25% | +2% | 1.19× | +0% |
| AutomationBench (Public) | 0.76 | +32% | −31% | −5% | 1.41× | −12% |
* HLE (w/o tools) 的 τ=1.02 是有意为正(即变慢):单轮任务上技能条目是净开销。