Diagrams — Module FTDD-08: Qwen3

Module: FTDD-08 — Qwen3 Diagram count: 4 Tool: Mermaid (primary). Each diagram validated in Mermaid Live Editor.

Diagram 1 — The Hybrid Problem Qwen3 Solves

Type: Comparison / tension Purpose: Why a dedicated reasoner is a production liability. A pure reasoner deliberates even on trivial queries; a pure fast assistant cannot reason. Qwen3 fuses both into one model. Reading the diagram: Top = the two extremes and their failure. Bottom = Qwen3's fusion: one model, two modes, switched by a thinking budget.

flowchart TD
  subgraph Ex["THE TENSION"]
    R["DEDICATED REASONER (e.g. R1)\nalways deliberates\nPRO: best on hard problems\nCON: latency tax on simple queries"]
    A["FAST ASSISTANT\nalways answers directly\nPRO: chatbot speed\nCON: cannot deliberate on hard problems"]
  end
  Naive["NAIVE FIX: two models + a router\n2x weights, 2x serving, router misroutes"]
  Q3["QWEN3: FUSION\none set of weights\nboth modes coexist\nswitched by a thinking budget"]

  R --> Naive
  A --> Naive
  Naive -.rejected.-> Q3
  Q3 --> Result["one deployment serves\nchatbot + reasoner\nno router, no second model"]

  style R fill:#14141f,stroke:rgba(255,255,255,0.12),color:#9494a0
  style A fill:#14141f,stroke:rgba(255,255,255,0.12),color:#9494a0
  style Naive fill:#08080c,stroke:rgba(255,255,255,0.08),color:#9494a0
  style Q3 fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#5eead4
  style Result fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8

Diagram 2 — The Four-Stage Qwen3 Post-Training Pipeline

Type: Sequential pipeline Purpose: Qwen3 mirrors R1 for two stages, then diverges. Stage 3 (fusion) is the novel contribution. Reading the diagram: Left to right. Stages 1 and 2 echo R1. Stage 3 is unique to Qwen3 — it merges thinking + non-thinking. Stage 4 aligns the whole.

flowchart LR
  S1["STAGE 1\nLong-CoT cold start\nseed reasoning format\nstructured <think>"]
  S2["STAGE 2\nReasoning RL (GRPO-style)\nverifiable rewards\nmath · code · logic"]
  S3["STAGE 3\nTHINKING-MODE FUSION\nmerge thinking + non-thinking\ninto one set of weights\n(the novel stage)"]
  S4["STAGE 4\nGeneral RL\nfull distribution\nkeeps non-thinking strong"]

  S1 -->|"anchors format"| S2
  S2 -->|"scales reasoning"| S3
  S3 -->|"one model, two modes"| S4
  S4 -->|"aligned hybrid"| Ship["Shippable Qwen3"]

  style S1 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style S2 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style S3 fill:#14141f,stroke:#5eead4,stroke-width:2px,color:#5eead4
  style S4 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style Ship fill:#08080c,stroke:#5eead4,stroke-width:1.5px,color:#5eead4

Diagram 3 — The Thinking Budget: Adaptive Compute

Type: Dial / spectrum Purpose: A single parameter controls how much the model deliberates. Low = chatbot speed; high = reasoner quality. Same weights. Reading the diagram: The dial runs from low (left) to high (right). Each position shows what a query produces. Same model, different output, controlled by one knob.

flowchart LR
  Low["LOW BUDGET\nshort/empty <think>\nfast direct answer\nchatbot speed\n\n'capital of France?'"]
  Mid["MID BUDGET\nbrief deliberation\nthen answer\nbalanced\n\n'write a regex for...'"]
  High["HIGH BUDGET\nlong self-correcting chain\nverified final answer\nreasoner quality\n\n'prove this converges'"]

  Low -.-> Mid
  Mid -.-> High

  Knob["THINKING BUDGET\none parameter\nlatency <-> quality dial\nsame weights, both modes"]

  Knob -.controls.-> Low
  Knob -.controls.-> High

  style Low fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style Mid fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style High fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8
  style Knob fill:#08080c,stroke:#5eead4,stroke-width:1.5px,color:#5eead4

Diagram 4 — R1 vs Qwen3: Which Reference When

Type: Decision comparison Purpose: The two canonical reasoning references. They prove different things. Choose based on what you are building. Reading the diagram: Left = R1. Right = Qwen3. The decision row at the bottom tells you which to study for which goal.

flowchart TD
  subgraph R1["R1 — DISTILLATION reference"]
    R1p["Teacher pipeline + SFT-only transfer\nProved reasoning emerges from RL\nProved traces transfer via SFT"]
  end
  subgraph Q3["QWEN3 — HYBRID reference"]
    Q3p["Fused thinking + non-thinking\nThinking-budget adaptive compute\nStrong non-thinking from 36T+ pretrain"]
  end

  R1 --> D1["GOAL: build a reasoning student\nby distilling from a teacher\n-> study R1"]
  Q3 --> D2["GOAL: deploy one model that\nserves chatbot + reasoner\n-> study Qwen3"]

  Both["STRONGEST TEAMS KNOW BOTH.\nQwen3's distillation drew on R1's recipe.\nField's budget conventions borrow from Qwen3."]

  D1 --> Both
  D2 --> Both

  style R1p fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8
  style Q3p fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8
  style D1 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style D2 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style Both fill:#08080c,stroke:rgba(94,234,212,0.4),stroke-dasharray:4 2,color:#5eead4

Validation notes

All four diagrams use the course design system colors: #14141f panel fill, #5eead4 accent for primary, rgba(94,234,212,0.5) for secondary borders, #e4e4e8 / #9494a0 for text.
Paste each into Mermaid Live Editor to render. All use stable Mermaid syntax (flowchart) supported in current Mermaid (v10.4+).
For the slide deck (artifact 03), these are rendered as static SVG/PNG captures from Mermaid Live, inlined into reveal.js.

# Diagrams — Module FTDD-08: Qwen3 **Module**: FTDD-08 — Qwen3 **Diagram count**: 4 **Tool**: Mermaid (primary). Each diagram validated in [Mermaid Live Editor](https://mermaid.live). --- ## Diagram 1 — The Hybrid Problem Qwen3 Solves **Type**: Comparison / tension **Purpose**: Why a dedicated reasoner is a production liability. A pure reasoner deliberates even on trivial queries; a pure fast assistant cannot reason. Qwen3 fuses both into one model. **Reading the diagram**: Top = the two extremes and their failure. Bottom = Qwen3's fusion: one model, two modes, switched by a thinking budget. ```mermaid flowchart TD subgraph Ex["THE TENSION"] R["DEDICATED REASONER (e.g. R1)\nalways deliberates\nPRO: best on hard problems\nCON: latency tax on simple queries"] A["FAST ASSISTANT\nalways answers directly\nPRO: chatbot speed\nCON: cannot deliberate on hard problems"] end Naive["NAIVE FIX: two models + a router\n2x weights, 2x serving, router misroutes"] Q3["QWEN3: FUSION\none set of weights\nboth modes coexist\nswitched by a thinking budget"] R --> Naive A --> Naive Naive -.rejected.-> Q3 Q3 --> Result["one deployment serves\nchatbot + reasoner\nno router, no second model"] style R fill:#14141f,stroke:rgba(255,255,255,0.12),color:#9494a0 style A fill:#14141f,stroke:rgba(255,255,255,0.12),color:#9494a0 style Naive fill:#08080c,stroke:rgba(255,255,255,0.08),color:#9494a0 style Q3 fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#5eead4 style Result fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8 ``` --- ## Diagram 2 — The Four-Stage Qwen3 Post-Training Pipeline **Type**: Sequential pipeline **Purpose**: Qwen3 mirrors R1 for two stages, then diverges. Stage 3 (fusion) is the novel contribution. **Reading the diagram**: Left to right. Stages 1 and 2 echo R1. Stage 3 is unique to Qwen3 — it merges thinking + non-thinking. Stage 4 aligns the whole. ```mermaid flowchart LR S1["STAGE 1\nLong-CoT cold start\nseed reasoning format\nstructured <think>"] S2["STAGE 2\nReasoning RL (GRPO-style)\nverifiable rewards\nmath · code · logic"] S3["STAGE 3\nTHINKING-MODE FUSION\nmerge thinking + non-thinking\ninto one set of weights\n(the novel stage)"] S4["STAGE 4\nGeneral RL\nfull distribution\nkeeps non-thinking strong"] S1 -->|"anchors format"| S2 S2 -->|"scales reasoning"| S3 S3 -->|"one model, two modes"| S4 S4 -->|"aligned hybrid"| Ship["Shippable Qwen3"] style S1 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style S2 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style S3 fill:#14141f,stroke:#5eead4,stroke-width:2px,color:#5eead4 style S4 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style Ship fill:#08080c,stroke:#5eead4,stroke-width:1.5px,color:#5eead4 ``` --- ## Diagram 3 — The Thinking Budget: Adaptive Compute **Type**: Dial / spectrum **Purpose**: A single parameter controls how much the model deliberates. Low = chatbot speed; high = reasoner quality. Same weights. **Reading the diagram**: The dial runs from low (left) to high (right). Each position shows what a query produces. Same model, different output, controlled by one knob. ```mermaid flowchart LR Low["LOW BUDGET\nshort/empty <think>\nfast direct answer\nchatbot speed\n\n'capital of France?'"] Mid["MID BUDGET\nbrief deliberation\nthen answer\nbalanced\n\n'write a regex for...'"] High["HIGH BUDGET\nlong self-correcting chain\nverified final answer\nreasoner quality\n\n'prove this converges'"] Low -.-> Mid Mid -.-> High Knob["THINKING BUDGET\none parameter\nlatency <-> quality dial\nsame weights, both modes"] Knob -.controls.-> Low Knob -.controls.-> High style Low fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style Mid fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style High fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8 style Knob fill:#08080c,stroke:#5eead4,stroke-width:1.5px,color:#5eead4 ``` --- ## Diagram 4 — R1 vs Qwen3: Which Reference When **Type**: Decision comparison **Purpose**: The two canonical reasoning references. They prove different things. Choose based on what you are building. **Reading the diagram**: Left = R1. Right = Qwen3. The decision row at the bottom tells you which to study for which goal. ```mermaid flowchart TD subgraph R1["R1 — DISTILLATION reference"] R1p["Teacher pipeline + SFT-only transfer\nProved reasoning emerges from RL\nProved traces transfer via SFT"] end subgraph Q3["QWEN3 — HYBRID reference"] Q3p["Fused thinking + non-thinking\nThinking-budget adaptive compute\nStrong non-thinking from 36T+ pretrain"] end R1 --> D1["GOAL: build a reasoning student\nby distilling from a teacher\n-> study R1"] Q3 --> D2["GOAL: deploy one model that\nserves chatbot + reasoner\n-> study Qwen3"] Both["STRONGEST TEAMS KNOW BOTH.\nQwen3's distillation drew on R1's recipe.\nField's budget conventions borrow from Qwen3."] D1 --> Both D2 --> Both style R1p fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8 style Q3p fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8 style D1 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style D2 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style Both fill:#08080c,stroke:rgba(94,234,212,0.4),stroke-dasharray:4 2,color:#5eead4 ``` --- ## Validation notes - All four diagrams use the course design system colors: `#14141f` panel fill, `#5eead4` accent for primary, `rgba(94,234,212,0.5)` for secondary borders, `#e4e4e8` / `#9494a0` for text. - Paste each into [Mermaid Live Editor](https://mermaid.live) to render. All use stable Mermaid syntax (`flowchart`) supported in current Mermaid (v10.4+). - For the slide deck (artifact 03), these are rendered as static SVG/PNG captures from Mermaid Live, inlined into reveal.js.