Signal-3.8-27B-Terse-Coder-FP8

Signal-3.8-27B-Terse-Coder

Merged weights of agentionai/Signal-3.8-27B with the Shockem/Qwen3.8-27b-Terse-Coder-LoRA adapter (round 8, rank-16 DPO) baked in. Signal is the adapter's training-lineage base — the adapter was trained partly on Signal's own traces, and the adapter card lists full-precision Signal as a recommended pairing, measuring a further ~40% reasoning-token cut on top of Signal's already-short traces with pass rate improving. This checkpoint is that pairing, pre-merged: no LoRA plumbing, no runtime adapter.

How it was made

• Merged in fp32: W + B @ A * (lora_alpha / r), with alpha 32 and r 16 (scale 2.0).
• Stored as bf16 using stochastic (unbiased) rounding with a fixed seed (0), so the merge is reproducible.
• MTP draft head and vision weights untouched — the adapter doesn't modify them, and Signal keeps the base model's MTP head, so speculative decoding remains available.
• All non-weight files (config, tokenizer, processor, index) copied from the Signal base; the chat template is Shockem/froggeric-terse-coder, the one the adapter was evaluated with. Signal ships the original Qwen3.8 template; serving this merge without the adapter's template changes agentic behavior.

Why stochastic rounding

The adapter's weight deltas are deliberately tiny (‖Δ‖/‖W‖ ≈ 4e-4–1e-3), below bf16's per-element resolution. The adapter card measures delta survival of only 31–61% under plain bf16 rounding vs 94–99.9% in fp16. Stochastic rounding is unbiased — each element is rounded up or down with probability weighted so its expected value equals the true merged value — so the delta is preserved on average while keeping the checkpoint at the base's bf16 dtype and size.

Expected behavior and caveats

• The effect compounds: Signal already produces ~52% fewer thinking tokens than stock Qwen3.8-27B [1], and the adapter cuts reasoning further on top of that.
• The adapter's author recommends runtime LoRA as the full-strength deployment form and measured a capability tax after merging on the stock Qwen base (70% → 60–62% on their held-out-40 after fp32-merge → fp16 → NVFP4 re-quant). This merge stores bf16 with no re-quant, so the tax should be smaller, but it is not zero, and no independent benchmarks have been run on this artifact.
• The adapter's measured Signal-base numbers used the NVFP4 quant of Signal; full-precision Signal is the listed recommended pairing, but treat the exact magnitude of the cut on this artifact as unmeasured.
• Do not load the Terse-Coder LoRA on top of this model. Double application over-shortens reasoning (63% pass with no_code failures in the adapter's testing).
• The adapter targets coding tasks with thinking enabled. It is a behavioral edit, not a knowledge edit — if a task needs long derivation, raise reasoning_effort as usual.
• Sampling: Signal's recommended stable settings are temperature 0.6, min-p 0.05, top-p 0.95, top-k 20 [1]. Note the conflict: the adapter card says to omit min_p entirely when serving with speculative decoding, because vLLM rejects it under spec decode. If you enable MTP spec decode, drop min-p.

How to use

Transformers

from transformers import AutoModelForImageTextToText, AutoProcessor
import torch

model_id = "vwdubb/Signal-3.8-27B-Terse-Coder-FP8"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

vLLM

vllm serve vwdubb/Signal-3.8-27B-Terse-Coder-FP8 \
  --dtype bfloat16 \
  --tensor-parallel-size 1 \
  --max-model-len 262144 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --port 8000

MTP speculative decoding (optional, head is included and untouched):

--speculative-config '{"method":"mtp","num_speculative_tokens":3}'

License

Both parent models are Apache 2.0, and this merge is released under the Apache License 2.0. Upstream copyright and license notices are retained.

Acknowledgements

• AgentionAI — Signal-3.8-27B, the base model.
• Shockem — the Terse-Coder LoRA and chat template.
• Qwen Team (Alibaba Cloud) — Qwen3.8-27B.