Modelpublic

jcbtc/Qwen3.8-27B-IU4-Kairic-Signal

sourceHugging Faceapache-2.0updated 13d agoView on Hugging Face

0likes391downloads

Model Card

[image]

Qwen3.8-27B · KAIRIC SIGNAL

AgentionAI's Signal tuning on CIRU's KAIRIC EDGE execution stack. A 27B local model for AMD Strix Halo, combining an IU4-accelerated body, a standard Q8 output head, and native multi-token prediction.

Signal is tuned for more direct answers and less unnecessary reasoning. This release brings that tuning into the existing KAIRIC build by replacing the vocabulary output layer. The remaining 865 GGUF tensors and all three acceleration files are unchanged.

Requires the custom KAIRIC ROCm runtime. This is not a stock llama.cpp, Ollama, or CUDA GGUF package.

What this release contains

Component	Detail
Tuning	Signal-3.8-27B by AgentionAI
Model body	CIRU KAIRIC EDGE hybrid quantization with IU4 acceleration files
Output head	Signal BF16 → standard Q8_0
Drafting	Original native MTP weights retained; launcher uses depth 4
Default sampling	Temperature 0.7 · top-p 0.95 · top-k 20 · min-p 0
Target hardware	AMD Strix Halo / gfx1151

v1.0 validation: source lineage, all-tensor integrity, and output-head conversion checks passed. This specific Signal build has not yet been inference-tested or benchmarked on AMD. Parent-model quality and speed results are not claimed for this release.

Download and run

Download the GGUF, all three .pfs files, and the runner into the same directory:

bash

hf download jcbtc/Qwen3.8-27B-IU4-Kairic-Signal --local-dir kairic-signal
cd kairic-signal
sha256sum -c SHA256SUMS

Build the pinned KAIRIC v1.2 runtime using BUILD.md, then launch:

bash

LLAMA_SERVER=/absolute/path/to/ROCmFPX/build-kairic/bin/llama-server \
ROCM_PATH=/opt/rocm \
PORT=8080 \
bash ./run-kairic-signal.sh

The runner defaults to a 32K context, thinking off, and Signal's sampling values. It disables greedy argmax shortcuts, draft backend sampling, prompt reuse, and the unsafe native M65 verifier. CONTEXT, HOST, PORT, and MODEL_ALIAS can be overridden through environment variables. Larger contexts need sufficient memory and separate validation.

bash

curl http://127.0.0.1:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"kairic-signal","messages":[{"role":"user","content":"Explain how a hash map works."}],"max_tokens":1024}'

The package is for text inference; a vision projector is not included.

Files

File	Size
`Qwen3.8-27B-IU4-Kairic-Signal.gguf`	15.48 GiB
`Qwen3.8-27B-Kairic-IU4-FFN.pfs`	7.99 GiB
`Qwen3.8-27B-Kairic-IU4-GDN.pfs`	1.88 GiB
`Qwen3.8-27B-Kairic-IU4-GDN-Output.pfs`	0.70 GiB

The .pfs files are prepacked execution views used by PromptForge. Keep all three alongside the GGUF. File size is not peak runtime memory.

Build verification

Only lm_head.weight differs between the pinned Signal and Qwen BF16 checkpoints. The first 17 shards have identical published SHA-256 values; every tensor in the final shard was compared directly. Tokenizer, configuration, and chat-template identities match.

The head was converted with the runtime's Q8_0 reference function. An independent implementation checked all 1,271,398,400 values with exact quantized-byte agreement. Every non-head GGUF tensor and all other payload bytes match the published KAIRIC donor. See build-report.json and SHA256SUMS.

These checks establish artifact integrity. They do not measure task quality, MTP acceptance, or generation speed.

Credits

—AgentionAI Signal-3.8-27B — output-layer tuning for concise responses.
—Qwen3.8-27B — base model, tokenizer, and MTP weights.
—CIRU KAIRIC EDGE — model body, IU4 execution views, and PromptForge integration.
—ROCmFPX, llama.cpp, and ROCm Composable Kernel — runtime and kernels.

Model weights: Apache 2.0. Runtime-derived launcher: MIT. This is a CIRU derivative of Signal, not an AgentionAI runtime release.