Agentionai Signal 3.8·27B poster: white SIGNAL wordmark over an orange particle trail on black

On our held-out general-prompt evaluation, Signal produces 57% fewer answer tokens and uses 52% fewer thinking tokens than the untouched Qwen3.8-27B, while matching or improving measured answer quality. The result is substantially faster end-to-end generation: on typical chat prompts, Signal can finish in less than half the wall time of the base model on the same hardware.

Signal gets there by being more direct rather than by truncating answers. It removes unnecessary preambles, excessive formatting, sign-offs, and explanatory narration while preserving the substance of the response. In thinking mode, it keeps the useful reasoning steps while spending fewer tokens describing the process.

Answer and thinking tokens vs. untouched Qwen3.8-27B, held-out general-prompt evaluation
Qwen3.8-27B, answer tokensbase
100 %
Signal, answer tokens57% fewer
43 %
Qwen3.8-27B, thinking tokensbase
100 %
Signal, thinking tokens52% fewer
48 %

Token counts normalized to the base model at 100% per category. Measured answer quality matched or improved over the base model across the same evaluation; see methodology below.

What changes, measured

We evaluated Signal against the untouched Qwen3.8-27B Q8_0 using the same server, sampling settings, and prompts, on an otherwise identical model file. All prompts in this evaluation were held out from tuning.

BuildFormatSizeTarget
Signal-Q4_K_M GGUF, mainline-compatible 16.4 GiB Any llama.cpp build, -ngl 99
Signal-Q5_K_M (recommended) GGUF, mainline-compatible 19.0 GiB Any llama.cpp build, -ngl 99
Signal-Q8_0 GGUF, mainline-compatible 28.7 GiB Any llama.cpp build, -ngl 99

The 57%/52% figures above are Q8_0 against Q8_0, the same reference point used for every comparison on this page. Each build is a drop-in replacement: same file layout, same flags, as the Qwen3.8-27B GGUF it replaces.

Browse all builds on HuggingFace ↗

Running Signal

Stock GGUF, no fork required. It's a drop-in replacement for llama.cpp setups that already run Qwen3.8-27B: same flags, same server, swap the model path.

llama-server -m Signal-Q5_K_M/Signal-3.8-27B-Q5_K_M.gguf \
  -ngl 99

Full setup instructions are in the repo's README on HuggingFace.