Documentation

Quickstart

Get started with TemuClaude in under 5 minutes.

Option 1 — Use the playground (no installation):

Open the playground → — ask anything, get a superior answer. No signup, no setup. 20 free queries/day.

Option 2 — API access:

curl -X POST https://temuclaude.com/api/chat \
  -H "Content-Type: application/json" \
  -d '{"messages": [{"role": "user", "content": "What is 9.9 vs 9.11?"}]}'

# Response: SSE stream with answer + orchestration metadata

Architecture

TemuClaude is one endpoint that orchestrates 8 AI models behind the scenes. When you ask a question, it:

1. Classifies your query (math, coding, reasoning, knowledge, creative, legal, health, agentic)
2. Estimates difficulty (trivial, medium, hard)
3. Routes to the best strategy:
   • Trivial (60%) → single cheap model (Hy3 Preview, free models)
   • Medium (30%) → specialist model (DeepSeek V4 Pro, GLM-5.2, Gemini 3 Flash)
   • Hard (10%) → full 10-layer fusion stack (3 models parallel + cross-review + aggregate + verify + QA + debate)
4. Returns one clean answer — orchestration is invisible but visible in the playground

Model Pool

TemuClaude uses 8 models, each with a specific role:

10-Layer Pipeline

For hard queries, TemuClaude runs up to 10 quality layers. Each is independently validated by published research:

1. Web Search — DuckDuckGo for knowledge queries (free, unlimited)
2. MoA 3-Layer Fusion — Propose → Cross-Review → Aggregate (65.1% AlpacaEval vs GPT-4o 57.5%)
3. Self-Consistency — PRM-weighted voting across N samples (+18.4% MATH)
4. Code Verification — Execute Python, verify output (ground truth)
5. Reflexion — Verbal reflection on failure, retry with context (91% HumanEval)
6. Self-QA Gate — 5-rubric score (LC, FC, CM, GA, CL), retry if < 8/10
7. Z3/SMT Verification — Logical consistency check with SMT solver
8. Budget Forcing — Append "Wait" to force longer reasoning (s1 paper)
9. Step-Level Verification — Verify each reasoning step independently (rStar-Math)
10. Frontier Fallback — Escalate to Claude Sonnet 5 (IQ 53) for hardest 2%

3-Tier Routing

TemuClaude routes queries by difficulty to minimize cost without sacrificing quality:

Difficulty estimation:
  Word count → 0-5 points
  Task type (math/reasoning/coding) → +2 points
  Keywords (explain, analyze, compare) → +1-2 points
  Total → 0-10 scale

Routing:
  Trivial (60% of queries) → 1 cheap model, 500 tokens
    Models: Hy3 Preview, free models ($0.00)
    
  Medium (30% of queries) → 1 specialist model, 4096 tokens
    Math/Coding → DeepSeek V4 Pro
    Knowledge → GLM-5.2
    Legal/Health → Gemini 3 Flash
    Creative → MiniMax M3
    
  Hard (10% of queries) → Full 10-layer fusion stack, 8192 tokens
    3 models in parallel → cross-review → aggregate
    + code verification + self-QA + reflexion + debate

MoA 3-Layer Fusion

For hard queries, TemuClaude uses Mixture-of-Agents (MoA) with 3 layers:

Layer 1: 3 models propose independently
  GLM-5.2 → response A
  DeepSeek V4 Pro → response B
  Gemini 3 Flash → response C

Layer 2: Cross-Review
  GLM-5.2 reviews B and C → improved A'
  DeepSeek reviews A and C → improved B'
  Gemini reviews A and B → improved C'

Layer 3: Aggregation
  GLM-5.2 synthesizes A' + B' + C' → final answer
  With structured analysis: consensus, contradictions, insights, blind spots

Self-Consistency

For math and reasoning questions, TemuClaude generates N samples and votes:

• Math: 3 samples at temperature 0.7
• Reasoning: 2 samples
• Each sample scored by Nemotron (PRM-weighted voting)
• Highest-scoring answer selected as final

Code Verification

For math and coding questions, TemuClaude generates Python code, executes it in a sandbox, and returns the verified output:

1. Model generates code to solve the problem
2. Code executed in sandboxed subprocess (no network, temp dir, timeout)
3. If output matches → answer is verified (ground truth)
4. If execution fails → trigger Reflexion: model reflects, retries

Self-QA Gate

Every answer is scored on 5 rubrics (USVA framework, extended from ATTS):

• LC — Logical Coherence (is the reasoning consistent?)
• FC — Factual Correctness (are the facts right?)
• CM — Completeness (does it address all parts?)
• GA — Goal Alignment (does it answer what was asked?)
• CL — Clarity (is it clear, concise, accessible?)

Score = average × 10. If score < 8/10, TemuClaude retries with Reflexion feedback — up to 2 times.

Reflexion

When the QA gate fails, TemuClaude generates a verbal reflection on what went wrong, then retries with that context:

1. QA gate fails (score < 8/10)
2. Nemotron generates reflection:
   "The answer has a logical error in step 3.
    The formula should use integration by parts,
    not substitution. Fix this and retry."
3. Original model retries with reflection as context
4. New answer is re-scored on all 5 rubrics
5. If still failing, retry once more (max 2 retries)

Budget Forcing

If the answer is suspiciously short for a hard problem, TemuClaude appends "Wait" to force the model to continue reasoning (s1 paper, arXiv:2501.19393):

Model answer: "The answer is 42."
  → Too short for a hard math problem
  → Append "Wait"
  → Model continues: "Let me verify... [longer reasoning]"

Z3 Logical Verification

For reasoning questions, TemuClaude extracts logical claims and checks them with a Z3 SMT solver:

• Extracts "if X then Y", "X implies Y" patterns
• Encodes as Z3 boolean constraints
• Checks satisfiability (no contradictions)
• If contradiction found → triggers multi-agent debate

Frontier Fallback

For the hardest 2% of queries where all other layers score low, TemuClaude escalates to Claude Sonnet 5 (IQ 53 — the highest available):

• Only triggers when QA score < 0.75 after all retries
• Query must match frontier criteria (prove, derive, theorem, system design, refactor)
• Frontier model gets the previous best answer as context
• Re-scored — if better, replaces the answer

Methodology

TemuClaude's benchmark scores are projected from:

1. Individual model scores from ArtificialAnalysis Intelligence Index
2. Proven improvement numbers from published research papers
3. Stack effects: each technique's proven gain applied to the baseline
4. Conservative estimates (lower bounds of published ranges)

Reproducibility

Full benchmark scripts are available on GitHub (MIT licensed). Run them yourself:

# Clone the repo
git clone https://github.com/notSaiful/temuclaude.git
cd temuclaude-research

# Install dependencies
pip install -r requirements.txt

# Run TemuClaude on HLE benchmark
python benchmarks/run_temuclaude.py --dataset hle --sample 100

# Run on GPQA Diamond
python benchmarks/run_temuclaude.py --dataset gpqa --sample 100

View full results on GitHub →

Projected Scores

TemuClaude vs frontier models across 8 benchmarks:

* Projected from research analysis. Live results pending ArtificialAnalysis verification.

REST API

TemuClaude exposes a single endpoint:

POST /api/chat

Request:
{
  "messages": [
    {"role": "user", "content": "What is 9.9 vs 9.11?"}
  ]
}

Response: SSE stream
  data: {"chunk": "9.9"}
  data: {"chunk": " is"}
  data: {"chunk": " larger"}
  ...
  data: {"orchestration": {...}}
  data: [DONE]

Streaming

TemuClaude uses Server-Sent Events (SSE) for streaming responses:

• POST request with messages array
• SSE response — chunks stream as data: {chunk}
• Orchestration data — sent after all chunks, before [DONE]
• [DONE] — signals stream end

Orchestration Data

The final SSE event contains orchestration metadata showing exactly how the answer was built:

{
  "orchestration": {
    "taskType": "math",
    "tier": "hard",
    "models": [
      {"name": "glm-5.2", "response": "...", "latency": 3.2, "correct": true},
      {"name": "deepseek-v4-pro", "response": "...", "latency": 8.3, "correct": true},
      {"name": "gemini-3-flash", "response": "...", "latency": 13.9, "correct": true}
    ],
    "aggregator": "glm-5.2",
    "consensus": 3,
    "qaScore": 8,
    "codeVerified": true,
    "totalLatency": "48.2",
    "cost": "$0.015",
    "techniques": ["moa-3-layer", "cross-review", "structured-aggregation",
                   "self-consistency", "prm-weighted-voting", "code-verification",
                   "reflexion", "usva-5-rubric-qa", "s1-budget-forcing"]}
  }
}

Authentication

Free tier: no authentication needed. Paid plans use Bearer token:

curl -X POST https://temuclaude.com/api/chat \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"messages": [{"role": "user", "content": "Hello"}]}'

Rate Limits

Error Codes

Media Orchestration

TemuClaude also orchestrates media generation — images, video, text-to-speech, and music. Same 10-stage pipeline (cache, intent, tier, parallel generation, judge, quality gate, reflexion, memory, return). Each media type has its own model pool, routing logic, and quality gates.

Image Generation

3-tier routing with best-of-N generation and LLM judge:

Unique routing for special cases:

• Vector/SVG output → Recraft V3
• Text in images → FLUX-2 Flex, GPT Image 2
• Extreme aspect ratios → Nano Banana 2
• Character consistency → Nano Banana 2, GPT Image 2
• Multilingual text → Seedream 4.5

# Generate an image
curl -X POST https://temuclaude.com/api/media/generate \
  -H "Content-Type: application/json" \
  -d '{"type": "image", "prompt": "a cat on a windowsill", "tier": "standard"}'

Video Generation

3-tier routing with async submit/poll pattern:

Unique routing for special cases:

• 4K video → Kling V3 Pro
• Dialogue/lip-sync → Google Veo 3.1
• Multi-input (images + clips + audio) → Seedance 2.0
• Long-form video → LTXV-2

Text-to-Speech

3-tier routing with voice selection and quality gating:

Music Generation

3-tier routing with lyrics support and quality gating:

Docker

Run TemuClaude with Docker (MIT licensed):

# Build the image
docker build -t temuclaude .

# Run on port 8000 with OpenRouter
docker run -p 8000:8000 -e OPENROUTER_API_KEY=your-key temuclaude

# Or with Ollama backend
docker run -p 8000:8000 -e OLLAMA_API_BASE=http://host.docker.internal:11434 temuclaude

Fly.io

Deploy to Fly.io (Mumbai region for lowest latency to India):

# Install flyctl
curl -L https://fly.io/install.sh | sh

# Login and create app
fly auth login
fly launch --image temuclaude

# Deploy
fly deploy

# Set secrets
fly secrets set OPENROUTER_API_KEY=your-key

Environment Variables