MODEL SPECIFICATION
Aether Ultra
Our flagship composite intelligence model. Hybrid Mixture of Experts (MoE) architecture natively equipped with a 2M token context window and intrinsic tool calling.
Performance Benchmarks
| Benchmark | Aether Ultra | GPT-4o | Claude 3.5 Sonnet | Gemini 1.5 Pro |
|---|---|---|---|---|
| MMLU (0-shot) | 89.4% | 88.7% | 88.3% | 85.9% |
| HumanEval (Pass@1) | 92.1% | 90.2% | 92.0% | 84.1% |
| MATH (0-shot) | 77.3% | 76.6% | 71.1% | 67.7% |
Artificial Analysis Benchmarks
| Metric | Aether Ultra |
|---|---|
| TTFT (Time To First Token) | 16.4ms average |
| Throughput | 188.5 Tokens/Second |
| Code Generation (HumanEval) | 92.4% pass@1 |
| Multi-Step Reasoning (MMLU-Pro) | 88.9% |
2M Token Context
Process massive codebases, hours of audio, and hundreds of research papers in a single prompt with near-perfect retrieval accuracy.
Native Tool Calling
Built-in autonomous agent capability that reasons about complex APIs, invokes external tools, and orchestrates intricate multi-step workflows.