以下部分引用 AI 总结，现阶段 AI 仍然有幻觉。请以内容的原文为准。
数据新鲜度群聊 07-14 ✓卖方 07-14 ✓Wrap 断档15天 ✕News 07-15 ✓X 断档12天 ✕页面生成 07-15
DeepSeek-V4
41 atoms · 跨 15 天 · 首见 2026-04-23 · 最近 2026-07-03
三色: 🟦 fact 28 · 🟥 take 13 · stance ▲15/▼3/◆11
来源: X 41
时态: fresh:31 · aging:10
标签: 好数字:16 · 好观点:13 · 好思考:10 · 好信源:7 · 好问题:1
叙事 (narrative) (13)
- 2026-06-27 · 模型性能评价（优于GLM-5.2）
still sucks compared to other open-source models like DeepSeek-V4 or Kimi models
tags: 好思考 · → daily
- 2026-06-11 · 上下文内存显存占用
FlashMemory直接能把显存占用压到 1.3GB!
tags: 好数字·好信源 · → daily
value: qty=1.3GB · date=2026-06-12
- 2026-04-24
[老化中] · 技术报告发布
DeepSeek-V4’s technical report is where that shift becomes concrete
tags: 好观点·好思考 · → daily
- 2026-05-22 · ALE-Bench 表现优于 Grok-4.3
Grok-4.3 is pretty terrible, basically worse than all the frontier chinese models like ... DeepSeek-V4
tags: 好观点 · → daily
- 2026-04-28
[老化中] · RL-max extent
it seems they didn't RL-max it much if at all outside some narrow STEM
tags: 好观点 · → daily
- 2026-04-27
[老化中] · LisanBench 表現
DeepSeek-V4 is looking rough on LisanBench
tags: 好观点 · → daily
- 2026-04-24
[老化中] · 技术报告对投资者有含义
I wrote about what it means for investors
tags: 好思考 · → daily
- 2026-04-23
[老化中] · 编码能力升级
DeepSeek-V4 is a massive upgrade for open-source coding
tags: 好观点 · → daily
- 2026-04-30
[老化中] · Vision 压缩: CSA 未压缩
but the question is why doesn't the CSA compress it down?
tags: 好问题 · → daily
value: direction=uncompressed
- 2026-04-24
[老化中] · 研究优化方向
the research direction is basically improve everything: Engram, better pre-training understanding, better data, better long horizon tasks, multimodalities.
tags: 好观点 · → daily
- 2026-04-23
[老化中] · SoTA状态
DeepSeek V4 is the new SoTA because
tags: 好观点 · → daily
- 2026-04-23
[老化中] · 编程竞赛表现
coding competitions.... making this the first time an open model has matched a closed model on this task
tags: 好观点 · → daily
value: comparison=首次匹配 · reference=闭源模型
- 2026-04-23
[老化中] · Agent能力
agent... slightly worse than frontier closed models
tags: 好观点 · → daily
value: comparison=稍逊 · reference=前沿闭源模型
反驳 (refute) (1)
- 2026-06-29 · does_not_force_Anthropic_to_lower_prices
DeepSeek-V4 does *not* force Anthropic to lower prices
tags: 好观点 · → daily
预测 (forecast) (1)
- 2026-07-03 · 完整版本预计在2026年7月中旬推出
likely when the full version of V4 comes out
tags: 好观点 · → daily
value: date=2026年7月中旬
事实 (fact) (26)
- 2026-05-02 ⭐ · prinzbench得分
DeepSeek-V4 (Pro) did not perform well on my benchmark. Its result (23/99) was comparable to those of older models like Grok 4 and Kimi K2 Thinking.
tags: 好数字·好观点·好信源 · → daily
value: qty=23/99 · date=未标注
- 2026-06-11 · 长文本评测准确率提升
官方测试在长文本评测集（如 LongBench-v2 等）上的准确率平均最终提升了 0.6%.
tags: 好数字·好信源 · → daily
value: qty=0.6% · date=2026-06-12
- 2026-04-23 · long context benchmark outperformance vs Gemini 3.1 Pro
Outperformed Gemini 3.1 Pro on long context benchmarks
tags: 好数字·好思考 · → daily
value: direction=outperformed
- 2026-05-22 · attention memory usage for 128x1M token sequences
530GB for 128 x 1 million token sequences with DeepSeek-V4s attention vs 23TB with good old GQA
tags: 好数字 · → daily
value: qty=530GB · date=current
- 2026-05-22 · 推理 effort 数量
only one reasoning effort ist used while iterating the solution
tags: 好信源 · → daily
- 2026-04-25 · 官方API降价
FYI: DeepSeek-V4 official API is now 75% off until May 5th!
tags: 好数字 · → daily
value: direction=75% off · date=直到5月5日
- 2026-07-03 · 将在腾讯云TokenHub平台提供服务
Tencent Cloud will offer DeepSeek-V4 models on its TokenHub platform run directly from DeepSeek's own network starting mid-July
tags: 好信源 · → daily
value: date=2026年7月中旬
- 2026-06-09 · context length支持500K
DeepSeek-V4 now runs 500K context
tags: 好数字 · → daily
value: qty=500K · date=不知道
- 2026-05-07 · 架构特性
DeepSeek-V4 架构特性：MLA（8 头 1 KV 头）、MoE（4 路由 + 1 共享专家）、Hyper-Connections（Sinkhorn 路由）、多 token 预测
tags: 好数字 · → daily
- 2026-05-07 · RoPE设计与挑战
DeepSeek-V4 leverages RoPE as its positional encoding scheme. However, its upgraded attention architecture raises two fundamental design challenges: • CSA & HCA adopt token compression, merging multiple tokens into one critical representation. The key question: inject RoPE before or after token compression? • The attention layer runs on MQA mode with shared KV representation. Directly applying RoPE rotation to KV will leak positional information into value matrix V — how to resolve this contamination?
tags: 好思考 · → daily
- 2026-05-07 · HCA中逆旋转优化
HCA adds an inverse rotation on final output O, converting absolute position expression back to standard relative position form.
tags: 好思考 · → daily
- 2026-05-07 · RoPE在CSA与HCA中的统一设计逻辑
DeepSeek-V4’s CSA and HCA both face KV cache compression and MQA shared KV troubles. The two modules follow identical RoPE processing logic; the following takes HCA as the typical case.
tags: 好思考 · → daily
- 2026-05-07 · RoPE置于压缩后的位置锚定规则
HCA selects the starting position of each 128-token segment as the position anchor for compressed K RoPE rotation.
tags: 好思考 · → daily
- 2026-05-01 · 使用MOPD进行全词汇蒸馏
DeepSeek-V4: trains a very large number (10+) of domain experts independently using domain-specific SFT and RL, then distills all of them into a single student. This paper interestingly uses full vocabulary distillation, which has very high memory overhead and is infrastructurally complex, instead of approximating KL with a single logit.
tags: 好信源 · → daily
- 2026-05-01 · 模型参数量
模型参数量: 1.60 trillion
tags: 好数字 · → daily
value: qty=1.60 trillion · date=2026-05-02
- 2026-04-30 · Vision 上下文窗口中的视觉 token 数
that's a bit above what you'd have with 14*14 patches (naively, 242 with margins)
tags: 好数字 · → daily
value: qty=242 with margins
- 2026-04-30 · Prefill 不享受 4x 折扣
the prefill doesn't get the 4x discount
tags: 好信源 · → daily
value: direction=not discounted
- 2026-04-25 · 训练使用 MXFP4
DeepSeekは V4の学習で MXFP4 (FP4) を活用ですが HW不明
tags: 好思考 · → daily
- 2026-04-25 · Memory Usage Reduction
Pro model uses 90 percent less memory than DeepSeek-V3.2
tags: 好数字 · → daily
value: qty=90%
- 2026-04-25 · Compute Reduction
Pro model uses nearly 75 percent less compute than DeepSeek-V3.2 for long tasks
tags: 好数字 · → daily
value: qty=75%
- 2026-04-24 · 模型发布
DeepSeek-V4リリースで Huaweiが速攻公開
tags: 好思考 · → daily
- 2026-04-23 · 训练数据量
32T tokens
tags: 好数字 · → daily
value: qty=32T tokens
- 2026-04-23 · 数学竞赛成绩
math... DeepSeek-V4 reaches a proof-perfect 120/120
tags: 好数字 · → daily
value: qty=120/120
- 2026-04-23 · compute efficiency improvement
Improved compute efficiency by around 4x
tags: 好数字 · → daily
value: qty=~4x
- 2026-04-23 · KV cache reduction
Reduced the KV cache by 10x
tags: 好数字 · → daily
value: qty=10x
- 2026-04-25 · State-of-the-Art Status
It now sets a new state-of-the-art for open models in reasoning, coding, and agentic performance
tags: 好观点 · → daily
时间轴 (近 20)
- 2026-07-03 ·
fact · 将在腾讯云TokenHub平台提供服务 · → daily
- 2026-07-03 ·
forecast · 完整版本预计在2026年7月中旬推出 · → daily
- 2026-06-29 ·
refute · does_not_force_Anthropic_to_lower_prices · → daily
- 2026-06-27 ·
narrative · 模型性能评价（优于GLM-5.2） · → daily
- 2026-06-11 ·
narrative · 上下文内存显存占用 · → daily
- 2026-06-11 ·
fact · 长文本评测准确率提升 · → daily
- 2026-06-09 ·
fact · context length支持500K · → daily
- 2026-05-22 ·
fact · attention memory usage for 128x1M token sequences · → daily
- 2026-05-22 ·
narrative · ALE-Bench 表现优于 Grok-4.3 · → daily
- 2026-05-22 ·
fact · 推理 effort 数量 · → daily
- 2026-05-07 ·
fact · 架构特性 · → daily
- 2026-05-07 ·
fact · RoPE设计与挑战 · → daily
- 2026-05-07 ·
fact · HCA中逆旋转优化 · → daily
- 2026-05-07 ·
fact · RoPE在CSA与HCA中的统一设计逻辑 · → daily
- 2026-05-07 ·
fact · RoPE置于压缩后的位置锚定规则 · → daily
- 2026-05-02 ·
fact · prinzbench得分 · → daily
- 2026-05-01 ·
fact · 使用MOPD进行全词汇蒸馏 · → daily
- 2026-05-01 ·
fact · 模型参数量 · → daily
- 2026-04-30 ·
fact · Vision 上下文窗口中的视觉 token 数 · → daily
- 2026-04-30 ·
narrative · Vision 压缩: CSA 未压缩 · → daily
← 实体目录 · 系统日志