Skip to main content
Back

AgentionAI's Signal Cuts Qwen Reasoning Tokens by 26% Without Losing Accuracy

A self-distilled fine-tune of Qwen3.8-Flash-Next cuts reasoning tokens by 26% and lifts speculative decoding acceptance from 69% to 90%.

  • AgentionAI released Signal-3.8-Flash-Next-GGUF, a terse self-distilled fine-tune of Qwen3.8-Flash-Next.
  • MATH-500: 26% fewer reasoning tokens, half the hesitation markers, 99.2% vs 97.5% accuracy.
  • Terminal-Bench 2.0: solves 10/10 tasks with 25% less wall time and 15% fewer tokens.
  • MTP draft head acceptance jumps from 69% to 90%, stacking speed on top of shorter outputs.
  • Tradeoff: on AIME 2025 at a 32k budget it answers 23/30 vs base's 25/30, with zero wrong answers.
  • Ships Q8_0, AP-Q4_K_XL and AP-IQ4_XS tiers with vision projector; drop-in for llama.cpp, Ollama, vLLM, LM Studio.

AgentionAI has released Signal GGUF, a self-distilled fine-tune of Qwen3.8-Flash-Next designed to produce shorter reasoning traces. The Hugging Face repository had recorded more than 17,000 downloads at publication.

Lower token counts can reduce latency, context consumption, and inference costs, especially when an agent makes repeated model calls. AgentionAI reports 26% fewer reasoning tokens on a 120-problem MATH-500 sample and 25% lower wall time on a 10-task Terminal-Bench 2.0 run. These are small, vendor-reported evaluations without cited independent replication.

Training out the hesitation loop

AgentionAI created Signal by training Qwen3.8-Flash-Next on its own answers generated under an instruction to respond directly. The released model no longer requires that instruction, and the company says it added no external training data or inference-time token bans.

The 177-billion-parameter mixture-of-experts backbone, vision support, and multi-token-prediction draft head retain their existing interfaces. A mixture-of-experts model routes each token through selected expert blocks, while the draft head predicts several upcoming tokens for speculative decoding. The unchanged architecture allows Signal to use the same serving stack as the base model, although the fine-tune can still change behavior and answer quality.

The release cites a quantization study that examined hesitation markers such as Wait, But, Alternatively, and maybe. The researchers found that uncertain tokens often trigger extra reasoning branches, sometimes after a model has already produced the correct answer. Suppressing a set of 50 such tokens during decoding reduced reasoning length by 12% to 23% across five quantized models without lowering reported accuracy.

A shorter trace on MATH-500

AgentionAI evaluated both models on 120 MATH-500 problems using the AP-Q4_K_XL quantization tier. Quantization compresses model weights to reduce memory use and can affect both speed and output quality, so the reported comparison applies most directly to this configuration.