EmpirioLabs Releases Aplomb 1: Open-Weights 5.3B Decision Model with 1M Context and Multimodal Input

EmpirioLabs has released Aplomb 1, an open-weights 5.3B decision model built on Qwen3.5-4B with the audio encoder from Qwen3-Omni-30B-Audio. The model supports 1M token context and accepts text, image, video, and audio inputs in a single request. It is designed to make decisions with probabilistic outputs for tool arguments, allowing agents to act on high-confidence calls while routing uncertain ones to larger models.

Performance and Benchmarks

On the Decision Index 0.2.1, Aplomb 1 scores 44.86, ranking #1 among 4B models on the published board. It achieves the top score among models up to 5.3B on 8 of 38 benchmarks, including MMLA, GPQA Diamond, and BBH.

API Capabilities and Latency

While the open weights are available, the 1M-token speed and decision features are currently exclusive to the EmpirioLabs API. Key API metrics include:

• Latency: Processes a 1M-token document in ~3 seconds (compared to ~111 seconds for a full read in standard mode). Short queries take ~15 ms model time (~111 ms end-to-end).
• Pricing: $0.00 per 1M input tokens, with free output and Zero Data Retention (ZDR) by default.
• Compatibility: Supports OpenAI, Anthropic, and Gemini formats alongside the native Decisions API.

Technical Details

The model expands the context window from 262K to 1M tokens and includes a custom decision head. It runs in bf16 on ~12 GB of GPU memory using the reference script. The weights are licensed under the EmpirioLabs License, free for research, evaluation, personal use, and internal use by companies with under $1M in annual revenue. Training data included public train splits of WinoGrande and ContractNLI.