The numbers

6.87 GiB

single GGUF file, full 27B

96.7%

HumanEval pass@1 (29/30)

92.9%

MMLU retention vs base

+3.85%

English perplexity cost

Benchmarks

Test	Caphlon-Model	Base quant	Result
English perplexity (EN40)	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%

Measured 25 October 2026, Tesla T4, llama.cpp, temperature 0, thinking disabled.

Hardware requirements

Setup	Minimum	Comfortable
GPU, full offload	8 GB VRAM (8K ctx)	12 GB VRAM (32K ctx)
CPU only	8 GB RAM (mmap)	16 GB RAM
Apple Silicon	16 GB unified	24 GB+
Disk	7 GB	—

Quickstart

llama-server

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

Measured on this build: 29.7 tok/s on file-edit workloads (2.34× the baseline 12.7), 12.4 tok/s on free-form prose. Send chat_template_kwargs: {"enable_thinking": false} or the model will spend your context on internal reasoning.

License and attribution

Repository contents: Apache-2.0. Weights are derived from Qwen 3.8 27B, distributed under Apache-2.0 — the upstream license governs them, permits commercial use, and requires that notices be retained.

"Caphlon" is a project mark. Apache-2.0 grants no trademark rights: you may redistribute the weights, but you may not present a modified build as official, endorsed, or authored by this project.