agentionai/Signal-3.8-27B

Signal 3.8 27B

This is Qwen3.8-27B that gets to the answer faster.

AgentionAI Signal is a minimally invasive fine-tune of Qwen3.8-27B designed for lower generation latency and better token efficiency. On our held-out general-prompt evaluation, Signal produces 57% fewer answer tokens and uses 52% fewer thinking tokens, while matching or improving the measured answer quality of the base model.

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 untouched Qwen3.8-27B 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.

Signal is trained by self-distillation: on Qwen3.8-27B's own answers, generated under an instruction to be direct that the released model no longer needs. No external data and no other model's outputs went into it, which is why it keeps the base model's knowledge and voice intact.

These are the full BF16 weights in Hugging Face format, a drop-in replacement for Qwen3.8-27B in transformers, vLLM, SGLang and any quantization pipeline. Quantized GGUF tiers from IQ4_XS to Q8_0 are in agentionai/Signal-3.8-27B-GGUF.

What changes, measured

We evaluated Signal against the untouched Qwen3.8-27B, both at Q8_0 in llama.cpp, using the same server, sampling settings, prompts, and otherwise identical model file. All prompts in these evaluations were held out from tuning.

	base Q8_0	Signal	change
general answers, median tokens	243	104	-57%
answers opening with a preamble ("Sure!", "Great question")	13%	0%	gone
answers with markdown headers	47%	18%	-62%
answers with bold	85%	52%	-39%
coding answers, median tokens	159	142	-11%
coding answers, p90 tokens	1026	914	-11%

Thinking mode, same prompts with reasoning on:

	base Q8_0	Signal	change
reasoning tokens, general prompts, median	153	74	-52%
reasoning tokens, coding prompts, median	225	166	-26%
reasoning tokens, GSM8K, median	119	81	-32%

Quality, exact match on GSM8K:

	base Q8_0	Signal
thinking off, 60 problems	98.3%	98.3%
thinking on, 40 problems	92.5%	95.0%

Shorter is not cheaper: no answer in the 100-prompt style set was cut off early (0 answers ending on a header or a colon, 0 unclosed code blocks), and no reasoning trace in 50 thinking-mode outputs looped or hit the token cap.

Faster with speculative decoding

Qwen3.8-27B carries a built-in multi-token-prediction draft head. Signal's answers are more predictable, so the drafter agrees with the model more often:

Draft acceptance and decode speed with --spec-type draft-mtp, both models Q8_0 on the same machine (Strix Halo, Vulkan), 200-token greedy runs for the fixed-draft rows:

prompt / draft length	base acceptance	Signal acceptance	decode speed vs base
prose, draft 3	39%	47%	+10%
prose, draft 4	35%	28%	-9%
structured output (JSON), draft 3	72%	94%	+20%
structured output (JSON), draft 4	66%	87%	+22%
chat prompts, sampled at 0.7, adaptive draft ≤4 (40 prompts)	57%	60%	—

Combined with the shorter answers, a typical chat reply finishes in well under half the wall time of the base model on the same hardware. Because of the specific finetuning there is no degredation over long context length.

What is in the repository

The complete Qwen3.8-27B checkpoint, 18 safetensors shards in BF16, with one tensor replaced: lm_head.weight. Every other tensor, the vision encoder, the projector, the MTP draft head, the tokenizer and the chat template are byte-identical to Qwen/Qwen3.8-27B. The head delta has a norm of 4.1% of the original head. Vision input works as in the base model.

Because only the output layer differs, any quantization recipe, LoRA, or serving setup that works on Qwen3.8-27B works on Signal unchanged.

Running

Thinking on and off both work; the chat template is the original Qwen3.8 template.

Sampling: temperature 0.7, top-p 0.95, top-k 20, min-p 0. Use sampling rather than greedy decoding. We saw a single loop at temperature 0.

transformers

from transformers import AutoProcessor, AutoModelForImageTextToText

model_id = "agentionai/Signal-3.8-27B"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(model_id, dtype="bfloat16", device_map="auto")

messages = [{"role": "user", "content": "Explain how a hash map works."}]
inputs = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True,
                                       return_dict=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=1024, do_sample=True, temperature=0.7, top_p=0.95, top_k=20)
print(processor.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Thinking is on by default. Pass enable_thinking=False to apply_chat_template to turn it off.

vLLM

vllm serve agentionai/Signal-3.8-27B --dtype bfloat16 --max-model-len 65536 --reasoning-parser qwen3

Then call the OpenAI-compatible endpoint with the sampling settings above. Send "chat_template_kwargs": {"enable_thinking": false} to turn thinking off per request.

llama.cpp

Use the prebuilt tiers in agentionai/Signal-3.8-27B-GGUF, which include the vision projector and the multi-token-prediction draft head, or convert this checkpoint with convert_hf_to_gguf.py.

Support AgentionAI

Signal3.8 is released freely. If it saves you compute or makes Qwen more useful, you can sponsor continued tuning, quantization and benchmarking on GitHub.