nika: v1
workflow:
  id: model-bench
  description: "One question → three local models → a measured comparison table"

model: ollama/qwen3.5:4b # the house default · contenders below pick their own seats

vars:
  question:
    type: string
    default: "Explain, in five lines, why a workflow should be audited before it runs."

tasks:
  ask_incumbent:
    infer:
      model: ollama/qwen2.5:14b # the quality bar · the seat to beat
      prompt: "${{ vars.question }}"
      max_tokens: 600

  ask_challenger:
    infer:
      model: ollama/llama3.2:3b # another family · same size class
      prompt: "${{ vars.question }}"
      max_tokens: 600

  ask_tiny:
    infer:
      model: ollama/qwen2.5:0.5b # the "is small enough?" probe
      prompt: "${{ vars.question }}"
      max_tokens: 600

  tabulate:
    with:
      ask_incumbent_duration_ms: ${{ tasks.ask_incumbent.duration_ms }}
      ask_incumbent: ${{ tasks.ask_incumbent.output }}
      ask_challenger_duration_ms: ${{ tasks.ask_challenger.duration_ms }}
      ask_challenger: ${{ tasks.ask_challenger.output }}
      ask_tiny_duration_ms: ${{ tasks.ask_tiny.duration_ms }}
      ask_tiny: ${{ tasks.ask_tiny.output }}
    invoke:
      tool: "nika:jq"
      args:
        input:
          - model: "ollama/qwen2.5:14b"
            ms: "${{ with.ask_incumbent_duration_ms }}"
            answer: "${{ with.ask_incumbent }}"
          - model: "ollama/llama3.2:3b"
            ms: "${{ with.ask_challenger_duration_ms }}"
            answer: "${{ with.ask_challenger }}"
          - model: "ollama/qwen2.5:0.5b"
            ms: "${{ with.ask_tiny_duration_ms }}"
            answer: "${{ with.ask_tiny }}"
        expression: >-
          "| model | latency | chars | answer (first 160) |\n|---|---|---|---|\n"
          + (map("| " + .model + " | "
               + (if .ms == null then "cached" else (.ms | tostring) + "ms" end) + " | "
               + (.answer | length | tostring) + " | "
               + (.answer | gsub("\n"; " ") | .[0:160]) + " |")
             | join("\n"))

  persist:
    with:
      tabulate: ${{ tasks.tabulate.output }}
    invoke:
      tool: "nika:write"
      args:
        path: "./bench/model-bench.md"
        create_dirs: true
        content: |
          # Model bench — same question, measured

          ${{ with.tabulate }}

          Facts only: latency from the run's clock, length in characters.
          The answers are side by side — the quality call is yours.

{"a":{"graph_format":2,"workflow":"model-bench","nodes":[{"id":"ask_incumbent","verb":"infer","model":"ollama/qwen2.5:14b","permits":[]},{"id":"ask_challenger","verb":"infer","model":"ollama/llama3.2:3b","permits":[]},{"id":"ask_tiny","verb":"infer","model":"ollama/qwen2.5:0.5b","permits":[]},{"id":"tabulate","verb":"invoke","tool":"nika:jq","permits":["tool: nika:jq"]},{"id":"persist","verb":"invoke","tool":"nika:write","permits":["fs.write: ./bench/model-bench.md","tool: nika:write"]}],"edges":[{"from":"ask_challenger","to":"tabulate","kind":"terminal-observation"},{"from":"ask_challenger","to":"tabulate","kind":"value"},{"from":"ask_incumbent","to":"tabulate","kind":"terminal-observation"},{"from":"ask_incumbent","to":"tabulate","kind":"value"},{"from":"ask_tiny","to":"tabulate","kind":"terminal-observation"},{"from":"ask_tiny","to":"tabulate","kind":"value"},{"from":"tabulate","to":"persist","kind":"value"}]},"engine":"0.105.0"}

showcase · tier 2 · builders

Model bench

The same question fanned to three local models in parallel; a jq fan-in tabulates measured latency and length: no judge model, facts only.

Pick your default model on a measured table, not vibes: zero keys, zero spend. · conformance-gated in nika-spec ↗ · re-proven at every push

The whole file

5 tasks · 3 waves · the plan falls out of the bindings, nothing is scheduled by hand.

t2-model-bench.nika.yamlsource

1nika: v12workflow:3  id: model-bench4  description: "One question → three local models → a measured comparison table"56model: ollama/qwen3.5:4b # the house default · contenders below pick their own seats78vars:9  question:10    type: string11    default: "Explain, in five lines, why a workflow should be audited before it runs."1213tasks:14  ask_incumbent:15    ◇infer:16      model: ollama/qwen2.5:14b # the quality bar · the seat to beat17      prompt: "${{ vars.question }}"18      max_tokens: 6001920  ask_challenger:21    ◇infer:22      model: ollama/llama3.2:3b # another family · same size class23      prompt: "${{ vars.question }}"24      max_tokens: 6002526  ask_tiny:27    ◇infer:28      model: ollama/qwen2.5:0.5b # the "is small enough?" probe29      prompt: "${{ vars.question }}"30      max_tokens: 6003132  tabulate:33    with:34      ask_incumbent_duration_ms: ${{ tasks.ask_incumbent.duration_ms }}35      ask_incumbent: ${{ tasks.ask_incumbent.output }}36      ask_challenger_duration_ms: ${{ tasks.ask_challenger.duration_ms }}37      ask_challenger: ${{ tasks.ask_challenger.output }}38      ask_tiny_duration_ms: ${{ tasks.ask_tiny.duration_ms }}39      ask_tiny: ${{ tasks.ask_tiny.output }}40    ◆invoke:41      tool: "nika:jq"42      args:43        input:44          - model: "ollama/qwen2.5:14b"45            ms: "${{ with.ask_incumbent_duration_ms }}"46            answer: "${{ with.ask_incumbent }}"47          - model: "ollama/llama3.2:3b"48            ms: "${{ with.ask_challenger_duration_ms }}"49            answer: "${{ with.ask_challenger }}"50          - model: "ollama/qwen2.5:0.5b"51            ms: "${{ with.ask_tiny_duration_ms }}"52            answer: "${{ with.ask_tiny }}"53        expression: >-54          "| model | latency | chars | answer (first 160) |\n|---|---|---|---|\n"55          + (map("| " + .model + " | "56               + (if .ms == null then "cached" else (.ms | tostring) + "ms" end) + " | "57               + (.answer | length | tostring) + " | "58               + (.answer | gsub("\n"; " ") | .[0:160]) + " |")59             | join("\n"))6061  persist:62    with:63      tabulate: ${{ tasks.tabulate.output }}64    ◆invoke:65      tool: "nika:write"66      args:67        path: "./bench/model-bench.md"68        create_dirs: true69        content: |70          # Model bench — same question, measured7172          ${{ with.tabulate }}7374          Facts only: latency from the run's clock, length in characters.75          The answers are side by side — the quality call is yours.76

nika inspect · engine 0.105.0 · vendored graph, never re-derived

7 declared edges · 3 waves · the plan falls out of the bindings

try itopen in the playground guidethe walkthrough · docs verbinfer verbinvoke toolnika:jq toolnika:write