Caphlon-Model v1.0.0 — 6.87 GiB Qwen 3.8 27B (GGUF)

Caphlon-Model v1.0.0 — 6.87 GiB build of Qwen 3.8 27B

A 6.87 GiB GGUF build of Qwen 3.8 27B. Vocabulary merge-table reduction
(100k target) + hybrid-aware pruning of layers 12–15. No distillation, full 27B
topology.

Measurements (2026-10-25, Tesla T4, llama.cpp, temp 0, thinking off)

Test	Caphlon-Model	Base quant	Result
English perplexity (wiki.test.raw, 512 ctx × 40 chunks)	6.6088	6.3638	+3.85%
HumanEval, 30 problems	29/30	30/30	96.7%
MMLU, 502-question subset	344/502	370/502	92.9% retained
File-edit throughput (--spec-type ngram-mod)	29.7 tok/s	12.7 tok/s	2.34×

Download

The 6.87 GiB file exceeds GitHub's per-asset limit, so it ships as 4 parts
(1800 MiB each) plus SHA256SUMS.

# 1. download all parts + checksum file into one directory, then:
cat part-aa part-ab part-ac part-ad > Caphlon-Model-IQ2_S.gguf

# 2. verify (expected: 8487b4e443834659b230e0bd529674ad65e068fc00380b996aaf11c873d63bc0)
shasum -a 256 Caphlon-Model-IQ2_S.gguf
# or, part-by-part:
shasum -a 256 -c SHA256SUMS

Expected size: 7375122272 bytes.

Run

llama-server -m Caphlon-Model-IQ2_S.gguf \
  -ngl 99 -c 32768 -np 1 \
  -b 4096 -ub 1024 --flash-attn on \
  --spec-type ngram-mod --spec-ngram-mod-n-match 24 \
  --spec-ngram-mod-n-min 48 --spec-ngram-mod-n-max 64 \
  --jinja --port 8080

Send chat_template_kwargs: {"enable_thinking": false} on chat requests.

Requirements

8 GB VRAM (full offload, 8K ctx) or 8 GB RAM (CPU/mmap); 12 GB VRAM or 16 GB
RAM comfortable. Disk: 7 GB.

License

Weights are derived from Qwen 3.8 27B (Apache-2.0); the upstream license
governs them. Repository contents: Apache-2.0. See
NOTICE —
"Caphlon" is a project mark and may not be used to present modified builds as
official or endorsed.