DeepSeek · 近 7 天
DeepSeek V4 Flash
0–100 分，越高代表社区使用体验越正面，不代表能力测试成绩。
更新于 2026-09-15 11:52 UTC · 2026-09-08 – 2026-09-15 UTC社区口碑分73.8/100170 条评论 · 近 7 天
截至 2026-09-15 11:52:03 UTC，DeepSeek 家族的 DeepSeek V4 Flash 在近 7 天有 170 条明确归属该版本的社区反馈，社区口碑分为 73.8/100。部分有效分类评分：通用文本 70.5/100 (n=63)；编程 60.1/100 (n=10)；推理 70.3/100 (n=14)。 评分方法与数据来源
查看长期变化与分析
模型生命周期
DeepSeek V4 Flash
发布于 2026-04-24 · 返回 DeepSeek
精确版本信号覆盖 2026-08-06–2026-09-14 · n=1,638
统计窗口：2026-08-18 00:00:00 至 2026-09-15 00:00:00 UTC · 评分方法：experience_score_v2。
趋势历史仍在收集中
当前窗口已达到要求；用于判定的独立分段已有 3/4 个。
生命周期每日更新 · 最新完整 UTC 日期 2026-09-14
- 固定 14 天基准
- 67.8 n=528
- 最近 28 天
- 68.7 n=837
- 较基准变化
- +0.9
- 90 天斜率
- – 分 / 30 天
14 天滚动体感趋势
每个每日更新的点汇总此前 14 个完整 UTC 日。虚线是固定发布基准；正式趋势判决仍使用互不重叠的独立 14 天分段。
用于趋势判定的独立 14 天分段已就绪 3/4 个。
哪些维度解释了变化
分别展示各维度内部的口碑变化与讨论构成变化。这是观察性贡献，不是因果证明。
主要负向口碑贡献：推理, 本地部署, 编程。
| 维度 | 基准 → 当前 | 维度变化 | 权重占比 | 口碑变化贡献 | 讨论构成贡献 |
|---|---|---|---|---|---|
| 推理 | 基准 → 当前78.0 → 64.3n=74 → 81 | 维度变化−13.8 | 权重占比15.2% → 9.9% | 口碑变化贡献−1.73 | 讨论构成贡献−0.20 |
| 本地部署 | 基准 → 当前64.3 → 61.7n=35 → 44 | 维度变化−2.6 | 权重占比5.9% → 5.1% | 口碑变化贡献−0.14 | 讨论构成贡献+0.04 |
| 编程 | 基准 → 当前69.9 → 69.8n=60 → 103 | 维度变化−0.0 | 权重占比11.5% → 12.8% | 口碑变化贡献−0.00 | 讨论构成贡献+0.03 |
| 通用文本 | 基准 → 当前63.7 → 64.4n=179 → 295 | 维度变化+0.7 | 权重占比32.5% → 33.5% | 口碑变化贡献+0.24 | 讨论构成贡献−0.04 |
| 速度与延迟 | 基准 → 当前67.8 → 73.0n=157 → 297 | 维度变化+5.3 | 权重占比30.3% → 36.6% | 口碑变化贡献+1.75 | 讨论构成贡献+0.19 |
| 图像/视觉 | 基准 → 当前样本不足n=0 → 3 | 维度变化– | 权重占比– | 口碑变化贡献– | 讨论构成贡献– |
| 角色扮演/创作 | 基准 → 当前样本不足n=9 → 7 | 维度变化– | 权重占比– | 口碑变化贡献– | 讨论构成贡献– |
| 安全与拒答 | 基准 → 当前样本不足n=15 → 9 | 维度变化– | 权重占比– | 口碑变化贡献– | 讨论构成贡献– |
| 视频生成 | 基准 → 当前样本不足n=0 → 0 | 维度变化– | 权重占比– | 口碑变化贡献– | 讨论构成贡献– |
样本不足维度的未解释贡献：+0.74 分。
生命周期分沿用主指数的体验信号权重，n≥30 后不做样本收缩。它衡量公开用户感知，不代表模型能力，也不能证明后端原因。
社区评论
精选近 7 天的不同意见，摘录条数不代表真实好评比例。
11 条精选摘录
展开原文
Deepseek flash barely ever thought much in think mode. today I realised why the thinking times reduced because of higher token per second
展开原文
It sucks for my use case
展开原文
Deep Seek flash is orchestrating a new feature using a custom mcp
展开原文
DSV4's implementation was broken beyond 90k context
展开原文
i get about 1.2k tps and 4k prefill with 2 of them with a kv cache pool in the native fp8 of 4.5m with my custom engine, with 8 slots you get about 80tps each so about 600 tps
展开原文
I mainly use Deepseek flash.
展开原文
made good progress with the 0731 release of DeepSeek Flash
展开原文
DeepSeek V4 flash is way better than GPT Luna ,I try them both
展开原文
For my Django site, yes. DSV4Flash handles it with ease.
展开原文
Quality I get is much better than what I paid anthropic/open ai last year
展开原文
DeepSeek Flash takes the spotlight
展开原文
Much better than DeepSeek-V4-Flash-0731, at least on my complicated codebase.
展开原文
I would personally use v4 flash 0731 than v4.1
展开原文
DS4 Flash was good until price hikes, and finding a provider that serves at high speed and without quantisation at the prior price is tricky.
展开原文
two dgx spark will run deepseek-v4-flash at maybe 30 tps. it is kinda slow, but it run stable.
展开原文
Flash isnt even remotely close to being as good for rp and worldbuilding like pro
展开原文
DeepSeek has to offload a ton to the CPU and it performed worse than Qwen in absolute terms and was a lot slower (not usable)
展开原文
DeepSeek flash feeled too boreing even pro I was breaking out this cycle but I am doomed to stay in
展开原文
DSv4 didn't even try to test its hypotheses with code to obtain empirical data
展开原文
pro v4 was great, this switch to flash has stripped a lot of nuance from it.. It's become very blunt, direct and even aggressive. The analyis and ideas are actually strong though, better than before, very good on that front, but socially it
展开原文
I have 4 Mi50s (32g) that runs v4 flash at q2 loading all weights in VRAM.
展开原文
I myself am content with being gpu poor and being able to run deepseek v4 flash at q2 and 11 tps for now
近 7 天暂时没有符合此筛选条件的精选评论。