agentionai/Signal-3.8-Flash-Next-GGUF

Signal 3.8 Flash Next

Qwen3.8-Flash-Next, tuned for speed and efficiency.

AgentionAI Signal is a minimally invasive fine-tune of Qwen3.8-Flash-Next for lower generation latency and better token efficiency. The 177B mixture-of-experts backbone, the vision encoder, and the multi-token-prediction draft head all work exactly as in the base model. On our held-out prompts Signal is more direct: it drops preambles, excessive formatting, sign-offs, and narration while keeping the substance, and 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-Flash-Next'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, so it keeps the base model's knowledge and voice intact.

What changes, measured

Speculative decoding gets faster. Flash-Next drafts with a separate multi-token-prediction head. Signal's answers are more predictable, so the drafter agrees more often. On coding output with --spec-type draft-mtp, both at this Q4_K_XL tier on the same machine:

	base	Signal
MTP draft acceptance	69%	90%

Higher acceptance means more tokens verified per forward pass, so decoding is faster on top of the shorter answers.

Agentic coding: same work, fewer tokens. We ran Signal and the untouched Qwen3.8-Flash-Next as the coding agent behind marshall on ten tasks from Terminal-Bench 2.0, real fix-a-repo and sysadmin tasks scored by each task's own test suite. Both at this AP-Q4_K_XL tier, same machine, MTP on, one trial per task.

	base	Signal
tasks solved	8 / 10	10 / 10
wall time, tasks both solved	—	25% less
output tokens, tasks both solved	—	15% fewer

Signal solved every task and used a quarter less wall time and 15% fewer tokens on the tasks both solved, biggest on the long ones (sanitize-git-repo 254 s vs 901 s, crack-7z-hash 336 s vs 707 s). Single trial per task, so treat the pass counts as indicative rather than a rigorous rate.

Where the tradeoff shows. Signal is tuned to be direct, which helps on agentic, tool-driven work: fewer tokens per step, faster loops, and it reaches the answer sooner. The flip side is problems that genuinely reward a long, exhaustive reasoning trail or brute-force enumeration, where the base model's willingness to keep grinding pays off and Signal's brevity can stop short. If your task is one of those, prefer the base model, or use multi prompt approaches.

Files

Agention Precision recipes: per-tensor quant types chosen for accuracy per gigabyte, built on the same recipes as the measured base Qwen3.8-Flash-Next-AP tiers.

tier	download	VRAM	experts (gate/up)
Q8_0	~180 GiB	~135 GiB	Q8_0 8.5 bpw — reference quality
AP-Q4_K_XL	94.3 GiB	~67 GiB	Q4_K 4.5 bpw
AP-IQ4_XS	84.4 GiB	~57 GiB	IQ3_S 3.44 bpw

VRAM is with the n-gram table offloaded to disk. More tiers to follow.

Vision: Signal keeps Qwen3.8-Flash-Next's image input. The projector mmproj-F16.gguf is at the repository root — download it alongside the tier.

Running

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

Sampling: temperature 0.7, top-p 0.95, top-k 20, min-p 0.

llama-server -hf agentionai/Signal-3.8-Flash-Next-GGUF:AP-Q4_K_XL \
  --jinja -ngl 999 -fa on \
  --temp 0.7 --top-p 0.95 --top-k 20 --min-p 0

Add the multi-token-prediction draft head for the acceptance above (needs a build with --spec-type draft-mtp); the draft is a separate GGUF, agentionai/Qwen3.8-Flash-Next-MTP-Q8_0-GGUF:

  --spec-type draft-mtp --model-draft Qwen3.8-Flash-Next-MTP-Q8_0.gguf \
  --spec-draft-adaptive on --spec-draft-n-max 4

Thinking is on by default. To turn it off per request, send "chat_template_kwargs": {"enable_thinking": false} with the chat completion.

🛠️ 7. Method and tooling

Signal is trained by self-distillation on the base model's own answers; no external data. The tiers are built with our own Rust tooling, agention-infer: gguf-pack for byte-exact, content-addressed tier builds and validation, gguf-info for inspection, and a per-tensor distortion solver behind the Agention Precision recipes. Every tier is measured against Signal's own BF16 on a held-out 2026 corpus before it ships.

Support AgentionAI

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