ThakiCloud

Enterprise Agent Automation Platform · Seoul, South Korea

One Paxis. Many Workflows. Any Cloud.

We turn digital work into AI-operated workflows. Paxis is the enterprise agent platform; Metis (inference) and Maxis (training) are the model layer that makes running those agents affordable, on Telox / Velox / Aegis infrastructure under one Signum control plane.

This organization publishes the model side of that stack — the checkpoints, recipes and benchmarks we actually serve, plus the evidence to reproduce the numbers on their cards.

Start here

Browse by release line, not by the flat model list — every public checkpoint belongs to exactly one line.

Serving quantization, by base model: Qwen3.8-27B Qwen3-Coder-30B-A3B Qwen3-30B-A3B Vision-Language

Which quantization should I pull?

Pick by the GPU you actually have, not by the biggest number in the filename.

How we report numbers

Model cards here state the sample size, the temperature, and what the measurement cannot tell you. Three habits we hold to:

We name the noise floor before we claim a win. Two checkpoints built by running the same quantization twice — same model, same settings, same calibration — diverged by 3.56pp on GSM8K. Anything smaller than that is unattributable, and we write it that way instead of taking credit.

We say when a metric cannot discriminate. If a benchmark's minimum detectable difference is larger than the gap we observed, the honest sentence is "we don't know," not "no regression."

We say when a judge is uncalibrated. Where an LLM judge has not been checked against human labels, the number is a judge-axis win rate — not human evaluation, and we label it as such.

Papers

Metric-Construction Coupling Inflates Measured Synthetic Dialect Recovery — arXiv:2610.00164. Synthetic Korean dialect data recovers 91.2% of the real-data gain on region identification but only 63.7% on comprehension — and when the scorer shares its marker inventory with the sentence builder, even that is overstated. Withholding marker types dropped dialectness recovery from 91.8% to 8.1%. Ships with KoDialectBench, 1,000 items across five regions.

Off-Target Effects of Response-Style Alignment in a Korean 27B Language Model — arXiv:2609.11291. Aligning Qwen3.8-27B to Korean response style also moved two behaviors nobody trained for: abstention on ambiguous social questions, and unprompted disclosure in securities guidance. Both moved through the model's emission policy — how often it answers and how much it says. The Human-KO cards should be read next to it.

SkillRet: A Large-Scale Benchmark for Skill Retrieval in LLM Agents — arXiv:2605.05726. The benchmark, index and embeddings behind the SkillRet line.

🌐 thakicloud.com · 📝 Tech blog · 💻 GitHub