imuximux
AI workbench for macOS

Models

Product model ids are stable contracts for IDE, Chat, CLI/TUI, and API. Prefer ids from /models — not raw upstream names.

Catalog (high level)

See /models for live cards, pricing pegs, and benchmarks.

  • auto / clavue-2.1-fast — volume chat, tools, caption rewrite (32K)
  • clavue / clavue-2.1 — writing, vibecoding, 64K loops
  • clavue-2.1-pro — long-horizon multi-agent
  • clavue-2.1-rev — review / PR / verify
  • clavue-2.1-search — hosted web search specialty
  • clavue-img — official image generation
  • clavue-music — official songs · POST /v1/audio/speech · one-liner 10s spark / TTT-ready 30s

Best practice (2026-08-20)

Thinking stays off unless the user asked for a hard problem. Do not send temperature=0. Tools are non-streaming. Music: one-liner 10s (hop spark); TTT-ready 30s — never 180 as one engine call.

  • Volume / rewrite / tools → auto or clavue-2.1-fast
  • Dev / writing / 64K → clavue-2.1
  • Thinking: omit or chat_template_kwargs.enable_thinking=false; hard tasks only: effort=medium
  • Tools: stream=false, 5–12 tools, return tool_calls + role=tool
  • Music: model=clavue-music; one-liner 10s (no local expand); TTT-ready 30s seed=11
  • Image: model=clavue-img, do not send steps/guidance

Recommended settings

Copy these. The API also locks sampling when you omit fields. Do not send temperature=0. Keep thinking off unless the user asked for a hard problem.

  • Daily / rewrite / tools: auto or clavue-2.1-fast · 0.7 / 0.8 / presence 1.5 · enable_thinking false
  • Writing / coding: clavue-2.1 · same instruct settings · 64K
  • Hard problem only: effort=medium · thinking sampling 1.0 / 0.95 · max_tokens ≥ 2048
  • Tools: stream=false, tool_choice=auto, 5–12 tools
  • Music: see /docs/music-generation — 30s, seed 11, voice="30"
  • Image: prompt only + aspect_ratio — no steps/guidance
{
  "model": "auto",
  "temperature": 0.7,
  "top_p": 0.8,
  "presence_penalty": 1.5,
  "max_tokens": 1024,
  "chat_template_kwargs": { "enable_thinking": false },
  "messages": [
    { "role": "system", "content": "Only output the result. No preamble." },
    { "role": "user", "content": "Classify as song, image, or chat. JSON only: {\"kind\":\"...\"}\nUser: morning pine light, a quiet female song, 30 seconds" }
  ]
}

Prompt demos

One task per request. Rewrite captions and lyrics in two calls. For songs, one-liners can go straight to POST /v1/audio/speech (10s). TTT-ready lyrics+caption skip spark.

# 1) Volume rewrite — caption (model: auto or clavue-2.1-fast)
system:
You are a song caption engineer. Expand the user's intent into an English Structured Caption.
Output only three headings: Global Metadata / Vocal Details / Arrangement.
250–450 words. No song title, no lyrics, no explanation.

user:
Duration: 30 seconds. Acoustic pop. Soft female, close-mic, no belting.
Theme: pine needles at a morning window, a quiet street.
No full drum kit, no celebrity names.

# 2) Writing / coding (model: clavue-2.1)
user:
Fix this function and explain the bug:
function median(a){a.sort();return a[a.length/2]}

# 3) Hard problem only — add effort=medium
user:
Compare two approaches for joining 90s + 90s audio on a chorus boundary. Give one recommendation and the command that verifies duration.

# 4) Image rewrite then /v1/images/generations (model: auto → clavue-img)
system:
Write a 60–120 word English director prompt: subject, wardrobe, scene, lens, light, one style. No masterpiece/8k, no negative list.

user:
半身人像，85mm，窗光，没有水印，竖构图。

How to pick

Start with auto or clavue-2.1-fast for volume. Move to clavue-2.1 for writing and coding. Pro / rev / search / img / music are specialty lanes.

Where to set the model

Every surface takes the same product model id string. Pick the id in the UI or pass it in API requests — do not paste raw upstream names.

  • Web Chat / chat.clavue.com — model picker in the composer header
  • imux Agent Chat — provider Official Clavue → model dropdown per session tab
  • Clavue TUI — /model or session settings inside clavue
  • API / SDK — "model": "clavue-2.1" (or auto, clavue-2.1-fast, clavue-img, …) in chat/completions

Fast vs premium pools

Membership tracks two pools. Fast SKUs (auto, *-fast) draw from the fast pool; flagship, pro, rev, search, and image SKUs draw from premium. Hitting one pool does not always empty the other — check Account → Usage.

Next