Signal-3.8-Flash-Next-HGN-v2

Plain, non-abliterated Signal HGN v2 conversion, published separately as Quat3rnion/Signal-3.8-Flash-Next-HGN-v2. No donor tensor replacement or composition was performed. This repository is not an alternate branch, folder or continuation of a donor build.

This is a baseline-plus-delta publication, not a self-contained snapshot. It contains the runtime files that differ byte-for-byte from the exact pinned Peonist baseline, plus necessary licenses, notices, provenance and assembly metadata. Use the assembler below to obtain the complete 14-file runtime package. Original GGUF and BF16 inputs are not uploaded.

Exact origins and conversion

• Language-model source: agentionai/Signal-3.8-Flash-Next-GGUF at immutable revision 4d30467d9f087cd88640762808c05d14131a329b, path Q8_0/Signal-3.8-Flash-Next-Q8_0.gguf.
• Source input: 188,225,032,736 bytes, SHA-256 98fb2414cdffccc16b9ff4222aae60015b1284cdea19a15232f389cae1a66cdd.
• Original-Qwen MTP/vision auxiliary tensor and asset origin: Qwen/Qwen3.8-Flash-Next at de4b8e4d43b917e7706784d8bb445c9af86a3540. These inputs were reconstructed from pinned original Qwen tensors and imported by the converter. The Signal publisher's statement that these auxiliaries are unchanged is provenance, not a byte comparison against Signal's quantized auxiliary tensors.
• The exact retained Signal GGUF chat template is included. The base tokenizer_config.json was prepared by replacing only its chat_template field with that template and reserializing JSON. Token-to-ID coordinates were checked. The final template/config must not be replaced with baseline versions just because other tokenizer assets match.
• Converter: bf16-to-hgn 0.1.0, profile v2, seed 0, threads 2, HT beam 128, HT and I4R backends HIP, Q4CP backend CPU. Executable SHA-256: bb64ff5dcc6e9ab4800c095cedec538aba1bfd08a456cfc8cb67278a2aa4c807.
• This is a conversion from the retained Q8_0 GGUF, not a claim of direct original-BF16 conversion or numerical equivalence to BF16. Imported HGN auxiliaries were decoded to canonical values and re-encoded; no stock HGN output was silently substituted.

The recorded full HIP decode verification completed successfully, attesting 1,658 source tensors, 2,349 output tensors and 118,785,466,780 payload bytes. That verifies the conversion's recorded numerical/provenance checks, not task quality.

conversion-manifest.json is the unmodified original schema-2 metadata, SHA-256 f1afa0920054838693c1dd0c14cc667383b3ecfcfd0b4b02c5c85d0d29fb7df4. Recursive inspection found no absolute filesystem-path strings, so no portable-path rewrite or derived manifest was needed. All tensor/layout identities remain those of the original conversion. Machine-local receipt files and queue settings are not published; portable allowlisted attestations and their original receipt hashes are in provenance/conversion.json.

Immutable baseline dependency

Baseline: peonist-ai/halogen-qwen3.8-flash-next

Required immutable baseline revision: 5cc17cea1a10b502b4a5db41d8d2b1a0f22dba26. All 16 baseline files were compared by exact size and SHA-256, rather than filename or assumed sidecar compatibility. Only these four runtime files match and are intentionally omitted here:

This omits 22,885,540 bytes of identical runtime payload. The remaining 10 runtime files total 118,789,817,058 bytes before the required publication metadata. The assembled 14-file package totals 118,812,702,598 bytes.

All three generated HGN files differ from the pinned baseline and are retained:

The main checkpoint is model.hgn; the converted MTP tensors are included in that main package. Keep the matching ngram and vision outputs. Do not substitute baseline main, ngram or vision files, or a different Signal template/config. assembly-manifest.json specifies all 14 output paths, source paths, exact sizes and SHA-256 hashes. Baseline sources always use the immutable revision above; the other ten sources use this publication's chosen immutable commit.

Reproducible assembly

Requirements: Linux with atomic renameat2 support, Python 3 and the official huggingface_hub package. Install the SDK in your preferred environment:

python3 -m pip install huggingface_hub

Choose an immutable commit of this plain repository. The following example resolves its current commit once, prints that commit for your reproduction record, downloads both metadata files at exactly that commit, then assembles:

import subprocess
import sys
from huggingface_hub import HfApi, hf_hub_download

repository = "Quat3rnion/Signal-3.8-Flash-Next-HGN-v2"
revision = HfApi().model_info(repository, revision="main").sha
print("Pinned plain publication:", revision)
assembler = hf_hub_download(repository, "assemble.py", revision=revision)
manifest = hf_hub_download(repository, "assembly-manifest.json", revision=revision)
subprocess.run([
    sys.executable, assembler,
    "--manifest", manifest,
    "--revision", revision,
    "--output-dir", "Signal-3.8-Flash-Next-HGN-v2",
], check=True)

To reproduce that exact release later, use the recorded 40-hex commit as revision instead of resolving main again. Use a new, nonexistent output directory. The assembler verifies every file's full SHA-256 and size before atomically installing the complete directory; it does not overwrite an existing output. It links only downloaded Hugging Face cache files. Add --copy only if you explicitly want to permit cross-filesystem copies; budget enough disk for that choice. Cached SDK authentication is used when available without printing a token. No original local conversion output is linked or mutated.

The manifest-driven assembler is reused byte-for-byte, SHA-256 64f04097077b2635d1badc0add78ee8abce29cc8a52d47ef509286d7e47cf35a. .gitattributes is separately handled Hub/SDK transport metadata, not an assembly file: pristine new-repository bytes are retained and the SDK deterministically extends them with sorted per-path LFS rules as needed. It is not claimed unchanged from the Peonist baseline.

Recorded plain-model performance, not quality

The completed run signal-plain-v2-hgn-halogen-0.18.5-depth7-gate095-098-20261015T125100Z used BetterBench 0.6.1 with corpus 1.0 and Halogen 0.18.5, sampled depth 7, gate 0.95,0.98. It measured 392 successful requests: 160 single-stream decode, 40 prefill, and 192 concurrency requests (48 each at levels 1, 2, 4 and 8). Warmups and a separate smoke request are excluded from those 392.

• Weighted elapsed decode throughput: 65.217879 tokens/s.
• Concurrency aggregate throughput at levels 1 / 2 / 4 / 8: 63.6743 / 69.7192 / 76.0788 / 76.0822 tokens/s.
• Sampled request settings: temperature 0.7, top-p 0.95, top-k 20, seed 42; 20 measured single-stream runs per category.
• Prefill target labels 2k–64k are workload targets, not measured token counts; actual median prompt lengths were 1,559 / 5,980 / 11,866.5 / 23,594 / 47,075.

The actual plain checkpoint identity was tied to the verified 14-file inventory, main size/inode and mappings of all three generated HGN outputs. An existing API routing alias was retained for request parity; the alias itself is not evidence of which weights ran. The summary is derived only from this plain run, not a donor benchmark. See provenance/benchmark.json for metrics, exact receipt hashes, runtime image pin and request configuration.

The measurement window was 2026-10-15 12:58:09–13:35:19 UTC, before the later embedding-NPU addition. Recorded native health lists decider-0.8b and qwen3.5-2b helpers. This is not a measurement of a three-NPU configuration or of subsequent production state.

These are one run's performance measurements, not universal device results: the official report did not populate GPU metadata. Elapsed throughput is not instantaneous speed inferred from bursty SSE gaps. Missing cached-token usage means unknown, not a proven cold cache. Tail p99 metrics are under-sampled. 48 decode samples reached their output length limit; none was recorded as truncated in reasoning. No IFEval or BFCL quality evaluation was completed, and no claim is made about accuracy, safety, retained quality, BF16 equivalence or superiority to another model.

License and notices

• LICENSE: exact pinned original Qwen Qwen Community License 1.0.
• LICENSE-SIGNAL: exact pinned Signal repository license, also Qwen Community License 1.0.
• NOTICE: origins, derivative-conversion notice and license reminders.

Both license files are 3,235 bytes, SHA-256 a0dc422560841fd68e06d974907f8b4c709bca44a67daad2b528437bdf676c08. The full terms control. This is not a blanket Apache-2.0 grant. In particular, the license has model-name display conditions for the stated large commercial thresholds and requires a separate Qwen license for commercial Model as a Service or AI Work Assistant businesses, subject to its internal-use exception. Qwen, Signal's publisher, the Peonist baseline/runtime authors and other contributors retain their respective rights; this derivative conversion is not an official endorsement.