Signal 3.8 27B

Signal 3.8 27B

Gives shorter, more direct answers than Qwen3.8-27B by using fewer answer and thinking tokens, so replies finish sooner. A light fine-tune that drops into the same tools as the base model.

Made by agentionai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Signal 3.8 27B is a light fine-tune of Qwen3.8-27B from AgentionAI, built to cut generation latency and token use. It drops preambles, extra formatting and sign-offs while keeping the substance of the answer. In the card's tests on the first release, median answer length fell from 243 tokens to 104.

Who it suits: It suits teams that already run Qwen3.8-27B and want faster replies, shorter outputs and lower token use without changing their tooling.

Industry | Who uses it | How they use it | Example
Software | Teams running chat assistants | Swap Signal in for the base model | A support chat whose replies finish in less time
Developer tools | Builders of coding assistants | Use it where shorter code answers help | Coding answers with a median 11 percent fewer tokens
Local AI | Hobbyists running llama.cpp | Load a GGUF tier and enable the draft head | JSON output with faster decoding than the base model
Research | People studying self-distillation | Examine a model tuned on its base model's own answers | Compare token counts before and after tuning

You can skip it if you like long, heavily formatted answers or do not run Qwen3.8-27B.

Where to use it:

• transformers, vLLM and SGLang, as a replacement for Qwen3.8-27B
• Any quantization pipeline, and llama.cpp through the separate GGUF repository
• Windows, macOS and Linux with Python, to download the files

Use case | Who | What they get back
Chat replies | Product teams | Direct answers with fewer preambles and headers
Coding help | Developers | Shorter code answers
Reasoning tasks | Developers | Thinking traces with fewer tokens

How to set it up: You need these first:

• Python on Windows, macOS or Linux
• Disk space for about 55.6 GB of BF16 weights, or about 13.9 GB for a 4-bit copy, which is an estimate
• A tool that runs Qwen3.8-27B, such as transformers, vLLM or SGLang

1. Install or update the Hugging Face download client.
2. Download this repository's files to your computer.
3. Load Signal in your tool where you would load Qwen3.8-27B.
4. Apply the card's sampling settings listed after the commands.

pip install -U huggingface_hub
hf download agentionai/Signal-3.8-27B

The card's settings for stable, short reasoning:

• Temperature: 0.6
• Min-p: 0.05, with top-p 0.95 and top-k 20
• KV cache: q8_0 or f16
• Multi-token-prediction draft head: on

How to use it: You use Signal like the base model: send chat prompts and read the replies. In llama.cpp, the card's speed tests turned the draft head on with one flag.

Tool | What it does
--spec-type draft-mtp | Turns on speculative decoding with the multi-token-prediction draft head
hf download | Fetches the repository's files to your computer

Example: You ask a general question that the base model would open with a cheerful preamble. Signal starts with the substance and uses fewer headers and bold marks. The card's median is 104 tokens against 243 for the base model.

For newcomers and experts: Signal is Qwen3.8-27B tuned to waste fewer words. A good first try is loading it where you use the base model and comparing the length and style of one answer.

For readers who know the field, it was trained by self-distillation on the base model's own answers, generated under a be-direct instruction the final model no longer needs. No outside data was used. Its answers are more predictable, so the draft head agrees more often. With three draft tokens on JSON, acceptance rose from 72 to 94 percent.

Load Signal in place of Qwen3.8-27B
Set temperature 0.6 and min-p 0.05
Turn on the draft head if you can
Send a chat prompt
Get a shorter, direct answer

Good to know:

• Settings: without min-p, long answers can fall into repetition on some draws, and a 4-bit value cache makes that certain.
• Re-download: the card says the weights were updated after the first release, so older downloads should be replaced.
• Numbers: the percentages come from the first release and the authors say they are being measured again on the updated weights.
• Tests: the card compares both models at Q8_0 in llama.cpp, with small sets such as 60 GSM8K problems with thinking off.
• Speed: with three draft tokens on JSON, decode speed rose 20 percent on one test machine, but four draft tokens on prose lost 9 percent.

Title: Median answer tokens on general prompts
Base model Q8_0: 243 tokens
Signal: 104 tokens

• License: Apache 2.0.
• Cost to run: none named. You need hardware for 27.8B parameters.
• Data it reaches: only the Hugging Face download of the files, as far as the source shows.
• Not for: none named in the text available here.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 27.8B
• Weights: about 55.6 GB in BF16; about 13.9 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 20,222
• Likes: 19
• License: apache-2.0
• Based on: Qwen/Qwen3.8-27B
• Quantized versions: 6
• Fine-tunes: 1
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, token-efficient, efficient-thinking, qwen3.8, qwen3.8-27b, terse, agentionai, conversational

Links

• Model: https://huggingface.co/agentionai/Signal-3.8-27B
• Files: https://huggingface.co/agentionai/Signal-3.8-27B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download agentionai/Signal-3.8-27B

Then follow the usage in its model card. Nothing in this kit runs by itself.