DeepSeek-V4-Pro

36 atoms · 跨 16 天 · 首见 2026-04-23 · 最近 2026-07-02

三色: 🟦 fact 30 · 🟥 take 6 · stance ▲8/▼3/◆8

来源: X 36

时态: fresh:36

标签: 好数字:28 · 好观点:6 · 好信源:4 · 好思考:2

叙事 (narrative) (3)

• 2026-05-26 · coding ability 接近 Opus 4.5
  I think Kimi-K2.6 and DeepSeek-V4-Pro are already close to Opus 4.5 in terms of coding ability
  tags: 好观点 · → daily
• 2026-05-26 · coding time horizons 落后几个月
  but more generally (including coding time horizons) they still lag a few months behind
  tags: 好观点 · → daily
  value: qty=几个月 · direction=lag
• 2026-07-02 · performance relative to scale
  V4-Pro is disappointing given its scale
  tags: 好观点 · → daily

预测 (forecast) (1)

• 2026-06-16 · inference cost per million tokens output with extended context
  even if you 10x the price per token to go up to a million context it's still just $2.5
  tags: 好思考 · → daily
  value: qty=$2.5

事实 (fact) (32)

• 2026-06-07 ⭐ · 完成全参数后训练
  a Shenzhen team completed full-parameter post-training of DeepSeek-V4-Pro on Huawei Ascend 910C chips.
  tags: 好数字·好信源 · → daily
  value: chip=华为昇腾910C · location=深圳团队
• 2026-06-20 · 排行榜排名
  第三名：DeepSeek-V4-Pro
  tags: 好数字·好信源 · → daily
  value: qty=No.3
• 2026-05-12 · 幻觉率排名
  DeepSeek-V4-Pro (1 in 92) and Claude Opus 4.7 (0 in 104) show the lowest hallucination rates on this task.
  tags: 好数字·好信源 · → daily
  value: rank=1 in 92 · context=AI文献综述质量基准
• 2026-06-27 · 推理成本降低幅度
  DeepSeek just made their inference ~5x cheaper at 50 TPS
  tags: 好数字 · → daily
  value: qty=5x · direction=cheaper
• 2026-06-23 · Terminal-Bench 2.0 Verified 得分
  On the Terminal-Bench 2.0 Verified subset, DeepSeek-V4-Pro achieves a score of approximately 72.0
  tags: 好数字 · → daily
  value: qty=72.0 · date=V4 paper
• 2026-06-16 · API输出定价
  DeepSeek-V4-Pro is $0.87 output at 1.6T@49B
  tags: 好数字 · → daily
  value: qty=$0.87 · date=2026-06-16
• 2026-06-16 · inference cost per million tokens output
  DeepSeek-V4-Pro costs $0.25 per million tokens output at 140 tokens/s
  tags: 好数字 · → daily
  value: qty=$0.25
• 2026-06-16 · inference token rate
  DeepSeek-V4-Pro … at 140 tokens/s
  tags: 好数字 · → daily
  value: qty=140 tokens/s
• 2026-04-26 · 输入缓存命中价格
  Pro模型在2026年5月5日前叠加2.5折限时优惠，DeepSeek-V4-Pro模型输入（缓存命中）降低至0.025元/百万Tokens
  tags: 好数字 · → daily
  value: qty=0.025元/百万Tokens · date=2026-04-26起至2026-05-05
• 2026-04-26 · 缓存命中输入价格
  DeepSeek-V4-Pro 缓存命中输入价格: $0.003625
  tags: 好数字 · → daily
  value: qty=$0.003625 · date=2026-04-26
• 2026-04-23 · MRCR任务表现
  DeepSeek-V4-Pro outperforms Gemini-3.1-Pro on the MRCR task
  tags: 好数字·好观点 · → daily
  value: comparison=优于 · reference=Gemini-3.1-Pro
• 2026-04-23 · 总参数
  DeepSeek-V4-Pro: 1.6T total
  tags: 好数字 · → daily
  value: qty=1.6T
• 2026-04-23 · 激活参数
  DeepSeek-V4-Pro: 49B active params
  tags: 好数字 · → daily
  value: qty=49B
• 2026-04-23 · 上下文长度
  DeepSeek-V4-Pro supports 1M context length
  tags: 好数字 · → daily
  value: qty=1M
• 2026-04-23 · 1024K处单Token FLOPs
  DeepSeek-V4-Pro at 1024K tokens FLOPs ~0.32T
  tags: 好数字 · → daily
  value: qty=0.32T
• 2026-04-23 · 1024K处累积KV Cache
  DeepSeek-V4-Pro at 1024K KV Cache ~5GB
  tags: 好数字 · → daily
  value: qty=5GB
• 2026-04-23 · Agentic Coding benchmark表现
  Open-source SOTA in Agentic Coding benchmarks
  tags: 好数字 · → daily
  value: direction=SOTA
• 2026-04-23 · 世界知识排名
  Leads all current open models, trailing only Gemini-3.1-Pro
  tags: 好数字 · → daily
  value: direction=第二
• 2026-04-23 · Math/STEM/Coding推理排名
  Beats all current open models in Math/STEM/Coding, rivaling top closed-source models
  tags: 好数字 · → daily
  value: direction=领先
• 2026-06-06 · completed full-parameter post-training on Huawei Ascend 910C cluster
  A Shenzhen team completed full-parameter post-training of DeepSeek-V4-Pro (1.6 trillion parameters) on a Huawei Ascend 910C cluster.
  tags: 好数字 · → daily
  value: qty=1.6 trillion parameters
• 2026-05-22 · API 定价
  Deepseek just made their 75% OFF discount on DeepSeek-V4-Pro API permanent. 1M input tokens (cache miss) at $0.435 and 1M output tokens at $0.87. The RMB pricing is 3 yuan and 6 yuan.
  tags: 好数字 · → daily
  value: qty=$0.435/1M input tokens (cache miss), $0.87/1M output tokens
• 2026-05-18 · 空间理解能力问题
  DeepSeek-V4-Pro 还有这方面的问题（无法保持元素在同一轴向上）
  tags: 好思考 · → daily
• 2026-04-30 · API 定价: 1M输入代币(缓存未命中)折扣价
  DeepSeek-V4-Pro 1M输入代币(缓存未命中)折扣价: $0.435
  tags: 好数字 · → daily
  value: qty=$0.435
• 2026-04-30 · API 定价: 1M输出代币折扣价
  DeepSeek-V4-Pro 1M输出代币折扣价: $0.87
  tags: 好数字 · → daily
  value: qty=$0.87
• 2026-04-30 · 免费使用限制
  with free V4-Pro ofc
  tags: 好信源 · → daily
  value: direction=free
• 2026-04-24 · 参数量
  DeepSeek-V4-Pro：1.6T
  tags: 好数字 · → daily
  value: qty=1.6T · date=2026-04-24
• 2026-04-24 · 激活参数量
  DeepSeek-V4-Pro：49B激活
  tags: 好数字 · → daily
  value: qty=49B · date=2026-04-24
• 2026-04-24 · 上下文长度
  DeepSeek-V4-Pro：上下文长度1M
  tags: 好数字 · → daily
  value: qty=1M · date=2026-04-24
• 2026-04-23 · 参数量
  DeepSeek-V4-Pro with 1.6T parameters (49B activated)
  tags: 好数字 · → daily
  value: qty=1.6T total, 49B activated
• 2026-04-23 · 参数规模
  DeepSeek-V4-Pro: 1.6T params (49B active)
  tags: 好数字 · → daily
  value: qty=1.6T params (49B active) · date=2026-04-24
• 2026-05-03 · Agent 核心代码实现机制程度
  DeepSeek-V4-Pro 这边是工程外围相对粗糙，但 Agent 核心是真的能跑
  tags: 好观点 · → daily
  value: direction=强
• 2026-05-03 · implemented GPT-5.5 architecture mechanisms in code
  DeepSeek-V4-Pro actually implemented them in code
  tags: 好观点 · → daily

时间轴 (近 20)

• 2026-07-02 · narrative · performance relative to scale · → daily
• 2026-06-27 · fact · 推理成本降低幅度 · → daily
• 2026-06-23 · fact · Terminal-Bench 2.0 Verified 得分 · → daily
• 2026-06-20 · fact · 排行榜排名 · → daily
• 2026-06-16 · fact · API输出定价 · → daily
• 2026-06-16 · fact · inference cost per million tokens output · → daily
• 2026-06-16 · fact · inference token rate · → daily
• 2026-06-16 · forecast · inference cost per million tokens output with extended conte · → daily
• 2026-06-07 · fact · 完成全参数后训练 · → daily
• 2026-06-06 · fact · completed full-parameter post-training on Huawei Ascend 910C · → daily
• 2026-05-26 · narrative · coding ability 接近 Opus 4.5 · → daily
• 2026-05-26 · narrative · coding time horizons 落后几个月 · → daily
• 2026-05-22 · fact · API 定价 · → daily
• 2026-05-18 · fact · 空间理解能力问题 · → daily
• 2026-05-12 · fact · 幻觉率排名 · → daily
• 2026-05-03 · fact · Agent 核心代码实现机制程度 · → daily
• 2026-05-03 · fact · implemented GPT-5.5 architecture mechanisms in code · → daily
• 2026-04-30 · fact · API 定价: 1M输入代币(缓存未命中)折扣价 · → daily
• 2026-04-30 · fact · API 定价: 1M输出代币折扣价 · → daily
• 2026-04-30 · fact · 免费使用限制 · → daily

← 实体目录 · 系统日志