Chimera

🎯 Problem & Solution

The Problem

No single AI model is best at everything. DeepSeek V4 excels at coding but misses nuance in creative work. Claude Sonnet 4 is a brilliant writer but weaker at structured reasoning. Kimi K2.7 leads code benchmarks but struggles with broad analysis. Picking one model means accepting its blind spots.

Today's "model routing" tools pick one model per task. Multi-model "fusion" products give identical prompts to every model — wasting tokens on duplicate work. Neither captures the power of collaborative intelligence: different models tackling different aspects of a problem, then merging their insights.

The Solution

Chimera is a dynamic deliberation engine. A single "dispatcher" model analyzes your prompt, designs a custom processing pipeline (DAG), assigns domain-specific subtasks to the best models for each, and an aggregator merges their work using dispatcher-written instructions. The result: an answer that's stronger than any single model could produce — at a fraction of the cost of running every model on every query.

🏗️ Architecture

┌─────────────────────────────────────────────────────────────┐
│                        CLIENT                          │
│              POST /v1/chat/completions                       │
│              { "model": "auto", "messages": [...] }          │
└────────────────────────┬────────────────────────────────────┘
                         ▼
┌─────────────────────────────────────────────────────────────┐
│                   DISPATCHER (1 call)                    │
│  Analyzes task → Designs DAG → Picks models by category     │
│  → Writes per-worker custom prompts → Writes merge recipe   │
└──────┬──────────────┬──────────────┬────────────────────────┘
       ▼              ▼              ▼
┌──────────────┐ ┌──────────────┐ ┌──────────────┐
│ Worker A      │ │ Worker B      │ │ Worker C      │
│ Code-focused │ │ Design-focus │ │ Analysis     │
│ (DeepSeek    │ │ (Claude      │ │ (GLM-5.2)    │
│  V4 Pro)     │ │  Sonnet 4)   │ │              │
└──────┬───────┘ └──────┬───────┘ └──────┬────────┘
       │                │                │
       └────────────────┼────────────────┘
                        ▼
┌─────────────────────────────────────────────────────────────┐
│                   AGGREGATOR (1 call)                    │
│  Merges heterogeneous results using dispatcher's recipe     │
│  → Structured output (JSON Schema) → Final answer           │
└─────────────────────────────────────────────────────────────┘

Dispatcher → Workers (parallel) → Aggregator. One dispatcher designs the entire run.

⚡ Core Features

🔌 Six Interfaces, One Engine

🌐 Supported Providers & Models

Via LiteLLM gateway with automatic provider routing. Direct API routes for DeepSeek and Z.AI (lower latency); OpenRouter for everything else.

💼 Use Cases

🛠️ Technology Stack

💵 Cost Model

Chimera is MIT-licensed open source — you pay only for the underlying model API calls. No per-request markup, no platform fees.

Actual costs vary by prompt length and model selection. Budget formations use DeepSeek V4 Flash (~$0.00014/1K tokens). Premium formations can include Claude Sonnet 4 or GPT-5.5. The dispatcher picks models based on category-weighted scoring with configurable price sensitivity (0.0 = pure quality, 1.0 = pure bang-for-buck).

🚀 Deployment

Docker (One Command)

git clone https://github.com/totalwindupflightsystems/chimera.git
cd chimera
export DEEPSEEK_KEY=sk-...   # only required credential
docker compose up -d
# → REST API on :8765, Web UI on :8765/web/

pip

pip install chimera-deliberation[full]
cp chimera.yaml.example chimera.yaml   # add API keys
chimera serve                            # start server

12-Factor Config

Every setting is overridable via environment variables or CLI flags. No YAML editing required for common deployments:

CHIMERA_PORT=8080 CHIMERA_DISPATCHER=deepseek/deepseek-v4-pro chimera serve

✅ Quality & Reliability

🗺️ Current Status & Roadmap