bomllm/turbo-thai-config

BOM LLM v12.0.0 — OWUI v0.13.0 + Fleet Audit + Stability

🇹🇭 Thai AI Model — Free for Thai people | GitHub

What's New in v12.0.0 (Nov 5, 2026)

• Grok 4.7 full fleet audit — 8/8 machines verified
• ComfyUI fleet — ALL v0.43.0 (7 instances synchronized)
• Kuma monitoring — v2.5.7 deployed on AM4 + NAS
• Chat residuals fixed — keepalive timeout + web-search config corrected
• LiteLLM incident 30 — emergency model restore completed (28 models OK)
• Stale job cleanup + memory defrag — 8 orphaned RUNNING jobs resolved
• GOD CONFIG v9.0 unchanged — 7/7 locked params verified
• BOM AI v4.0.0 — training pipeline + ship release
• Full changelog: https://github.com/siamcafe/bomllm/tree/v12.0.0

What's New in v11.0.0 (Nov 4, 2026)

• Open WebUI v0.13.0 — upstream security fixes, 18 functions verified on the shared BOMLLM+BOMAI instance
• C2 performance survey + fix — external TTFB median 0.093 s (AM5 10-shot), SSH brute-force guard persisted to boot, LiteLLM DB-load keys verified at GOD CONFIG v9.0 target, pollers audited
• GOD BLUEPRINT v5.2 + HOW-TO v4.2 — new ops recipes (guard maintenance, LiteLLM DB checks, poller policy, external TTFB measurement, headless QC)
• GOD CONFIG v9.0 unchanged — 7/7 locked params verified live
• BOM AI v3.0.0 — shared-instance sync release
• Full changelog: https://github.com/siamcafe/bomllm/tree/v11.0.0

---

What's New in v10.0.0 — OWUI v0.12.5 Fleet Update (Nov 2, 2026)

• Open WebUI v0.12.4 -> v0.12.5 (18 mods intact, 12/12 QC PASS)
• Theme Designer Pro 1.8.0 -> 1.19.0
• Fleet cleanup: old containers removed, Docker pruned 6.4GB
• docker-compose.yml updated to match production versions
• GOD CONFIG v7.0 unchanged
• Incident #26 documented (C2 polling flood, fix dispatched)
• Full changelog: https://github.com/siamcafe/bomllm/tree/v10.0.0

What's New in v9.0.0 — Project Closure (Nov 1, 2026)

• Fleet 9/9 verified, pipeline E2E PASS
• 40 IRON safety rules enforced
• GOD CONFIG v7.0 locked fleet-wide
• FINAL BLUEPRINT + HOW-TO published
• SSH incident #24 fixed
• Benchmark: Thai gate 20/20, Han leak 0/5, speed 18.5 tok/s
• Pinned: OWUI v0.12.4, LiteLLM v1.106.4, Ollama 0.40.10-0.40.13, ComfyUI v0.43.0
• Full changelog: https://github.com/siamcafe/bomllm/tree/v9.0.0

What's New — v0.8.1 (Oct 31, 2026)

• GOD CONFIG v7.0 documentation — locked parameter table with reasons (below)
• QC system on AM4 — OCR / Whisper / ASR / Gateway, 10/10 PASS
• Grok deep survey — 30 findings across 7/7 machines, fleet fixes applied
• pp=0 verified — presence_penalty 0 confirmed on 15 models fleet-wide
• Full changelog: https://github.com/siamcafe/bomllm/releases/tag/v0.8.1

GOD CONFIG v7.0

Locked generation parameters for all production models:

Parameter	Value	Reason
temperature	0.6	Thai quality gate — ≥0.7 fails
top_p	0.95	Balanced sampling
top_k	20	Focused vocabulary
repeat_penalty	1.0	No repetition bias
presence_penalty	0.0	⛔ CRITICAL — 1.5 causes Han leak
num_predict	4096	Max generation length
num_ctx	8192	Context window

⛔ DO NOT change presence_penalty from 0.0 — values >0 cause Chinese character leakage in Thai responses.

What's New — GOD CONFIG v7.0 (Oct 30, 2026)

• GOD CONFIG v7.0 — 9 parameter corrections, fleet-wide standardization
• Ollama 0.40.13 on UA/AM4 (Mac remains 0.40.10)
• LiteLLM v1.106.4 — v1.107.0 supply-chain compromise blocked
• Fleet cleanup — 210.43 GiB reclaimed across UA/AM4/Mac/C2/NAS
• BOM AI v0.3.2 — verified PASS

What's New in v7 (Oct 27, 2026)

• GOD CONFIG v7.0 — 9 production config changes
• Open WebUI v0.12.4 — upgraded from v0.11.3
• LiteLLM v1.106.3 — security patched (v1.107.0 supply-chain compromise blocked)
• Ollama 0.40.10 — upgraded from 0.40.1
• Qdrant v1.20.0 — upgraded from 1.19

Model Details

• Base: DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU
• Parameters: 27.78B (27,781,081,984)
• Layers: 1,204
• Quantization: nvfp4 (MLX) / Q4_K_M (GGUF)
• Size: 21.58 GB (MLX) / 18.50 GB (GGUF)

Benchmark Results

Test	Score	Details
Thai Gate	20/20 = 1.00	10 questions, all 2/2, thai chars 292-1064
Han Leak	0/5 = 0.00	Zero Chinese character leakage
Street Test	3/3 PASS	thai>50, han<3, bytes>200
Speed	18.5 tok/s	Range 17-20, Apple M4 Pro MLX

License

MIT / Apache-2.0