DeepSeek-V4-Pro, Kimi K2.6, GLM-5.1 : 三款LLM接受Vibe Coding长上下文的考验

@startuml
skinparam backgroundColor #FEFEFE
skinparam defaultTextAlignment center
skinparam wrapWidth 220

title 测试环境 — Opencode 会话 × 3 LLMs
left to right direction

package "🖥️ 开发者终端" #E8F5E9 {
  rectangle "AGENT.adoc\n(规则，待办事项)" as AG
  rectangle "PROMPT_REPRISE\n(任务)" as PR
  rectangle "INDEX.adoc\n(路线图)" as IDX
}

package "☁️ Ollama 云" #BBDEFB {
  rectangle "DeepSeek-V4-Pro\n1.6T / CSA+HCA" as DV4
  rectangle "Kimi K2.6\n1T / MLA" as KIMI
  rectangle "GLM-5.1\n744B / MLA+DSA" as GLMG
}

package "⚙️ 代码库 Gradle" #FFF9C4 {
  folder "buildSrc/" {
    file "codebase.kt"
    file "readme.kt"
    file "site.kt"
    file "slider.kt"
    file "snapshot.kt"
  }
  file "build.gradle.kts\n(947 行)"
  file "embeds.yml"
}

AG --> DV4 : Sessions 1-8 ✅
AG --> KIMI : 第9-10节 ⚠️
AG --> GLMG : 第11-13节 ⚠️

DV4 --> "代码库" : "TDD, 重构,\n快照"
KIMI --> "代码库" : "降级\n自 60K 令牌"

note bottom of GLMG
比 Kimi 更好
但延迟更高
DSA 稳健性较差
比 CSA+HCA
end note

@enduml

@startuml
skinparam backgroundColor #FEFEFE
skinparam defaultTextAlignment center

title 比较雷达 — 3 LLMs 在辅助软件开发中
legend right
  |= Couleur |= Modèle |
  | <#FF5252> | Kimi K2.6 |
  | <#FFC107> | GLM-5.1 |
  | <#4CAF50> | DeepSeek-V4-Pro |
endlegend

rectangle " " as space #FFFFFF

rectangle "一致性\n80K+ tokens" as coh #F5F5F5
rectangle "质量\ncode" as qual #F5F5F5
rectangle "推理\narchitectural" as arch #F5F5F5
rectangle "抵抗\nhallucinations" as hall #F5F5F5
rectangle "速度\n(开发体验)" as speed #F5F5F5
rectangle "知识\nGradle/Kotlin" as kg #F5F5F5
rectangle "纠正率\ncorrection" as corr #F5F5F5

coh --> qual
qual --> arch
arch --> hall
hall --> speed
speed --> kg
kg --> corr

note top of coh
  Kimi      ████░░░░░░ 4/10
  GLM       ██████░░░░ 6/10
  DeepSeek  █████████░ 9/10
end note

note top of qual
Kimi      ███████░░░ 7/10 (低于60K)
GLM       ████████░░ 8/10
DeepSeek  █████████░ 9/10
end note

note top of arch
  Kimi      █████░░░░░ 5/10
  GLM       ███████░░░ 7/10
  DeepSeek  █████████░ 9/10
end note

note top of hall
  Kimi      ██████░░░░ 6/10 → ██░░░░░░░░ 2/10 (> 60K)
  GLM       ███████░░░ 7/10
  DeepSeek  █████████░ 9/10
end note

note top of speed
  Kimi      █████████░ 9/10
  GLM       ██████░░░░ 6/10
  DeepSeek  █████░░░░░ 5/10
end note

note top of kg
  Kimi      ███████░░░ 7/10
  GLM       ███████░░░ 7/10
  DeepSeek  █████████░ 9/10
end note

note top of corr
Kimi      ██░░░░░░░░ 2/10 (大量更正)
GLM       █████░░░░░ 5/10
DeepSeek  ██████████ 10/10 (几乎没有需要修改的地方)
end note

@enduml

@startuml
skinparam backgroundColor #FEFEFE
skinparam defaultTextAlignment center
skinparam wrapWidth 260

title 决策树 — 为辅助开发选择自己的 LLM
start

:初始上下文\n> 20K 标记 ?;

if (Oui) then
  :注意力\n混合或多层架构？;

  if (CSA+HCA) then
    :DeepSeek-V4-Pro
    Fenetre stable 128K
    Ideal pour gouvernance EAGER;
    stop
  elseif (MLA + DSA) then
    :GLM-5.1
    Correct jusqu'a ~120K
    Latence elevee;
    stop
  else (MLA seul)
    :Kimi K2.6 (sessions courtes)
    ou eviter pour long contexte
    Chercher un modele avec
    sparse attention;
    stop
  endif
else (Contexte < 20K)
  :N'importe lequel fonctionne.
  Privilegier la latence.;
  stop
endif

@enduml