agentionai/Signal-3.8-27B

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 10, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

agentionai/Signal-3.8-27B is a 27 billion parameter language model based on Qwen3.8-27B, fine-tuned by AgentionAI for reduced generation latency and improved token efficiency. It achieves significantly fewer answer and thinking tokens while maintaining or enhancing answer quality, making it substantially faster for end-to-end generation. This model is optimized to produce more direct responses by removing unnecessary preambles and excessive formatting, making it ideal for applications requiring concise and efficient AI outputs.

Loading preview...

Overview

agentionai/Signal-3.8-27B is a minimally invasive fine-tune of the Qwen3.8-27B model, developed by AgentionAI. Its primary goal is to achieve lower generation latency and better token efficiency compared to its base model. Through self-distillation, Signal 3.8-27B learns to be more direct in its responses, eliminating unnecessary preambles, excessive formatting, and verbose narration without compromising the substance or quality of the answer.

Key Capabilities

  • Reduced Token Usage: Produces 57% fewer answer tokens and uses 52% fewer thinking tokens on general prompts compared to the base Qwen3.8-27B.
  • Faster Generation: Achieves less than half the wall time for end-to-end generation on typical chat prompts.
  • Concise Responses: Eliminates preambles, excessive formatting, and sign-offs, leading to more direct and actionable outputs.
  • Maintained Quality: Matches or improves measured answer quality, including a 95.0% exact match on GSM8K with thinking on, compared to 92.5% for the base model.
  • Enhanced Speculative Decoding: Shows increased draft acceptance rates (e.g., 94% for structured output with draft 3) and faster decode speeds (up to +22%) when used with Qwen3.8-27B's built-in multi-token-prediction draft head.
  • Full Compatibility: As only the lm_head.weight tensor is replaced, it is a drop-in replacement for Qwen3.8-27B and compatible with existing quantization recipes, LoRA, and serving setups.

Good for

  • Applications requiring fast, efficient, and concise AI responses.
  • Scenarios where reduced computational cost per generation is critical.
  • Use cases benefiting from direct answers without conversational fluff.
  • Developers already using Qwen3.8-27B who want to improve inference speed and token efficiency without changing their existing infrastructure.

Finetunes
2 models