Skip to content
siamcafePublic

About

Self-hosted Thai AI infrastructure — Mac Mini M4 Pro running TURBO-Fable 27B at 18.5 tok/s for $0.54/day. Fleet of 9 machines, GOD CONFIG v7.0, 40 IRON safety rules. Production-grade Open WebUI + LiteLLM + Ollama MLX stack.

Topics

Resources

Contributing

Security policy

Stars

48 stars

Watchers

0 watching

Forks

Repository files navigation

BOM LLM — Self-Hosted Thai AI Infrastructure

BOM LLM v18.0.0

BOM LLM

BOM LLM v18.0.0 — Fleet Sync & Documentation Release: GOD CONFIG v12, GOD BLUEPRINT v8, LiteLLM v1.108.8 LTS, BOM AI Pro GGUF, Gateway rc3, IRON 57+ Rules

License: MIT Fleet Status Pipeline GOD CONFIG Open WebUI IRON Rules LiteLLM Thai NLP Self Hosted HuggingFace

The fleet today (Nov 14, 2026): 8 machines · RAM 1,200+ GB · 8 GPUs · Tailscale-meshed · 34 incidents survived · 57+ IRON rules.

I replaced $50/day of cloud LLM bills with a Mac Mini on my desk running at $0.54/day measured electricity. It now does inference, coding, image generation, voice synthesis, music, and live streaming — all self-hosted.

One Mac Mini M4 Pro 48GB runs TURBO-Fable (27.78B nvfp4 MLX) at 17–20 tok/s (avg 18.5, verified 2026-10-25). A MiMo-V2.6 Q4 (9B) runs alongside at 37 tok/s for local coding drafts. GLM-5.3 on cloud handles production coding with 71/72 benchmark score. An i9 with 2× RTX 5060 Ti renders images (ComfyUI + FLUX), Thai voice (SiangTTS), and music (YuE2 + ACE-Step). Everything meshes over Tailscale; public HTTPS via Cloudflare Tunnel; every API key-gated.

This is not a demo. This is the exact stack that answers my customers every day.

Architecture

Full fleet architecture and policies live in BLUEPRINT.md (GOD BLUEPRINT v8.0 — full fleet survey, priority matrix, evidence contract + v8.1 addendum: Pro GGUF tier, gateway rc3 stream resume, web search default OFF) and HOWTO.md (GOD HOW-TO v5.0 — install/deploy steps, Ollama pull commands, daily sweeps, incident recipes §3.8-§3.17). Locked generation parameters: GOD_CONFIG.md (v12.0 — keep_alive "24h" everywhere, never -1).

What's New in v18.0.0 — 📦 Fleet Sync & Documentation Release (Nov 14, 2026)

BOM LLM v18 poster

  • GOD CONFIG v12 (hub art#60163) + GOD BLUEPRINT v8 (hub art#57374) fully synced into this repo — keep_alive "24h" rule fleet-wide (NEVER -1)
  • Grok deep scan: 8 machines / 462 files (art#60726) — 8 findings fixed (art#60745): bomai-safe keep_alive, typhoon25-content, admission_lanes, chart_reason, intent filter port, UA swap, UA LiteLLM restart policy, CB1 zombies
  • typhoon25-content permanently killed from LiteLLM DB (art#60808); LiteLLM + OWUI restarted with all fixes live (art#60753)
  • LiteLLM v1.108.5 → v1.108.8 LTS applied on C2/UA/AM4 · ComfyUI v0.45.0 fleet · Ollama 0.40.16 unification target (fleet survey art#60843)
  • BOM AI v2.1.8 — Pro tier now GGUF Q4_K_M on all platforms (was MLX-only)
  • Gateway rc3 stream resume — Han-cut rate 0/272 (was 13.97% on v2.1.6)
  • Web search default OFF — 3 layers fixed (model preset, admin default, builtin tools)
  • Workflow guide published (art#60694, 11 chapters) · IRON 38 → ABSOLUTE · incidents #33/#34 closed · 7 CAT nginx CVEs handed to project web
  • Fleet specs: RAM 1,200+ GB · 8 GPUs · 8 machines · Mac Studio M5 Ultra 512GB on order (Jan-Feb 2027)

BOM AI Model Tiers (v2.1.8 — Pro GGUF)

Tier Model Format Size Ollama Tag
Light 4B GGUF ~3 GB siamcafe/bom-ai:light-v2.1.8
Pro 709-L 27B (DavidAU) GGUF Q4_K_M 16.5 GB + 518 MB mmproj siamcafe/bom-ai:pro-v2.1.8

keep_alive policy: "24h" on every route — models stay warm; -1 is FORBIDDEN (incident #33).

What's New in v17.0.0 — 🛡️ Infrastructure & Reliability Update (Nov 8, 2026)

Benchmark v17

  • Open WebUI v0.13.0 → v0.13.2 — coding agents, saved side chats, comments, notifications, plain-HTTP fix
  • LiteLLM v1.106.5 → v1.108.3 LTS — cosign signature verified on upgrade, Postgres restart fix
  • Valkey v9.1.2 → v9.1.3 — heap overflow security fix; Docker DNS alias persisted via recreate script + compose
  • Ollama i9 v0.40.13 → v0.40.15 · Cloudflared Mac v2026.8.2 → v2026.10.7
  • Contabo C2 live-migrated — disk latency 133ms → 4ms (90% improvement)
  • bom_price_pipe — blocking time.sleep → async asyncio.sleep (event loop unblocked)
  • bom-writer — num_ctx 16384 pin added to LiteLLM route
  • Deep fleet audit — 8 machines, 25 findings, 3 critical fixes applied
  • Chat 500 fixed — Valkey alias root cause resolved (incident #31)
  • 2 config files hardened 664 → 600 · LiteLLM cosign verification on every upgrade
  • GOD CONFIG v11.0 + GOD BLUEPRINT unchanged — hero banner v17 + benchmark card via ComfyUI FLUX (dark blue & gold)

What's New in v16.0.0 — ⚡ Atom Fleet Optimization Complete (Nov 7, 2026)

  • GOD CONFIG v11.0 (hub art#56686) + GOD BLUEPRINT v7.0 (hub art#56687) — atom-level fleet survey 225 files, 23 optimization actions implemented
  • Premium hero banner v16 — ultra-premium dark+gold cinematic design via ComfyUI DreamShaper
  • Atom survey complete — 8 machines, 225 files, 54600 lines scanned by Grok 4.7 (art#56608)
  • K3 optimization — 23 actions designed by Kimi K3-256k (art#56647)
  • zcode Wave 1+2 — 11/11 actions DONE: price router v2, UA memory guard, C2 OWUI memcap, AM4 log rotation, CB1 zombie watch, live-router model fix (art#56649)
  • zcode Wave 3 — AM4 postgres rebound, GOD v11/v7 published (art#56683)
  • Price Router v2 — narrowed intent regex (7 symbols only), 2s timeout, honest HTTP poll source label
  • live-router model fix — qwen38-chat → qwen38-uni, /healthz verified 200
  • LiteLLM v1.106.5 — fleet-wide reconciled pin (v1.107.0 remains BANNED)
  • Fleet config backup — NAS git tag config-snapshot-2026-11-06 (1928 files, 8.83 MB)
  • Fleet cleanup — 4.8 GB reclaimed (docker prune + i9 temp)
  • BOM AI v2.0.3 — Han Fix HF-1.0 integration, bom_runtime 2584 lines
  • IRON Rules 44 — added IRON 44 (Grok dry-run before zcode implement)
  • Netdata — 5 nodes verified (AM4/C2/Mac/CAT/UA), NAS removed
  • ComfyUI release banner — DreamShaper 8, premium dark+gold theme

What's New in v13.0.0 — 🚀 Mega Fleet Optimization (Nov 6, 2026)

  • K3 fleet optimization executed — 35 actions from the K3 blueprint (hub art#56543) across all 8 machines in 13 phases, spec art#56548
  • Live price pipeline — BOM Assistant now answers XAUUSD (and 6 more symbols) from the owner MT5 API in real time (bom_price_pipe + bom_price_router + bom_price_tool on C2 OWUI); the $3,247 hallucination is fixed (answers ~live $4,08x)
  • GOD CONFIG v10.0 (hub art#56564) + GOD BLUEPRINT v6.0 (hub art#56565) — v9 data retained, v10 deltas marked
  • Bind hardening — UA Netdata + AM4 Netdata republished to loopback/tailnet/LAN only; NAS old node-exporter stopped after Prometheus scrape-free proof
  • Ollama 0.40.14 alignment — Mac CLI aligned (server already 0.40.14, qwen38-uni 7/7); AM4 + i9 upgrades staged as scripts, deferred by FX28 burn + 03:00-06:00 ICT gates
  • Cleanup — C2 exited containers reclaimed (~3 GB layers) with rollback archives; AM4 kern.log 235 MB rotated under a dedicated rule; agent-hub-v2_prev retired
  • Known deferred — UA LiteLLM bind (03:00-06:00 ICT window), C2 OWUI explicit bind (container has no Compose labels → owner migration), Contabo1 chromium zombies (parent is n8n — owner window), i9 ESU decision, NAS swap refresh (needs root)

What's New in v12.0.0 — 🔍 Fleet Audit + Stability (Nov 5, 2026)

  • Grok 4.7 full fleet audit — 8/8 machines verified (art#56080)
  • 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 — 8 orphaned RUNNING jobs resolved
  • Memory defrag — consolidated project memory for next phase
  • Core versions unchanged: OWUI v0.13.0, LiteLLM v1.106.5, GOD CONFIG v9.0

What's New in v11.0.0 — ⚡ C2 Performance Update (Nov 4, 2026)

  • Open WebUI v0.13.0 — upstream security fixes, 3 new migrations, 18 functions verified operational on the shared BOMLLM+BOMAI instance
  • C2 performance survey + fix — external TTFB measured properly (AM5 10-shot median 0.093 s), SSH brute-force guard persisted to boot, LiteLLM DB-load keys verified at GOD CONFIG v9.0 target, pollers audited
  • SSH guard policy — /etc/c2-ssh-guard.list + idempotent systemd unit; NEVER-BLOCK ranges enforced in the script
  • GOD BLUEPRINT v5.2 + HOW-TO v4.2 — new §3.13-§3.17 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-thai-chat, bomai-safe)
  • Residuals closed — charming_gauss removed, squid tailscale-rebind automated, /static/bomai/ hero banners serving 200
  • IRON Rules 43 — IRON 42 (title format) + IRON 43 (structured reply) added
  • Fleet 8/8 verified + Stress 7/8 PASS — 8 tailscale peers; T3 web-search passed on retry at 106.6 s (180 s window)

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, 14 overlays rebased, 12/12 QC PASS
  • Theme Designer Pro 1.8.0 → 1.19.0
  • docker-compose.yml pinned to production — OWUI v0.12.5, LiteLLM v1.106.5, Qdrant v1.20.0
  • Fleet cleanup — old containers removed, orphan skill IDs cleared, Docker pruned 6.4 GB
  • GOD CONFIG v7.0 NO CHANGE — locked params verified across 15 models
  • Hero banner v10 — images/hero_banner_v10.png

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

  • PROJECT CLOSURE 10/10 — all 14 findings resolved, fleet 8/8 verified, pipeline E2E PASS
  • IRON 40 Rules — 3 new safety rules (IRON 38/39/40) after SSH incident #24
  • SSH Incident #24 Fixed — PasswordAuthentication restored on all 8 fleet machines
  • Grok Deep Root Cause Scan — 8 machines × 9 areas, 3 root causes traced (hardware/config, not software)
  • FINAL BLUEPRINT — 51KB, 15 sections, comprehensive fleet architecture reference (art#54961)
  • FINAL HOW-TO — 32KB, 14 sections, complete operational manual (art#54962)
  • Mega Closure Report — 14/14 findings verified and closed (art#55017)
  • Pinned versions verified — LiteLLM v1.106.5, OWUI v0.12.4, Ollama 0.40.10-0.40.13, ComfyUI v0.43.0

What's New in v0.8.1 — 🛡️ GOD CONFIG Docs + QC System (Oct 31, 2026)

  • GOD_CONFIG.md — full GOD CONFIG v7.0 reference (locked params, per-model compliance, protected deviations)
  • QC System on AM4 — OCR / Whisper / ASR / Gateway, deployed Oct 31, 10/10 PASS
  • Grok deep survey — 30 findings across 7/7 machines; fleet fixes F001–F010 applied
  • K3 LFM3 lightweight model blueprint — Typhoon2-3B rank 1, lfm3:1.7b rank 2 (deployed)
  • Pinned versions — LiteLLM v1.106.5 (⛔ v1.107.0 COMPROMISED), ComfyUI v0.43.0 fleet-wide

What's New in v0.8 — 🔧 Mega Tuning & Fleet Optimization (Oct 30, 2026)

  • GOD CONFIG v7.0 — 9 parameter corrections across 5 model configurations, all presence_penalty=0 enforced
  • Fleet cleanup — 210.43 GB reclaimed across UA, AM4, Mac, C2, NAS
  • Ollama 0.40.13 — UA and AM4 upgraded (Mac stays 0.40.10 for production stability)
  • bomai-safe deployed — content rewrite model on UA and AM4
  • Dead route cleanup — 9 live + 3 staged removals across LiteLLM instances
  • Docker cleanup — 15 unused images removed (19.33 GB reclaimed)
  • Auto-start fixes — i9 llamacpp-gpu1 Automatic, NAS nas-hub unless-stopped
  • Model inventory — 10 orphan model paths + 1 stale Ollama tag removed
  • Fleet expanded — 8 machines documented (added UA, AM4, CAT)

Previous: v7 — 🔒 Security & Fleet Update (Oct 27, 2026)

— 🔒 Security & Fleet Update (Oct 27, 2026)

  • GOD CONFIG v7.0 — 9 production config changes (Ollama 0.40.10, router truth, rollback containers)
  • Open WebUI v0.12.4 — upgraded from v0.11.3 (Python 3.12.15, Pillow 13)
  • LiteLLM v1.106.3 — security patched; v1.107.0 supply-chain compromise blocked
  • Ollama 0.40.10 — upgraded from 0.40.1 (versioned LaunchDaemon pattern)
  • Qdrant v1.20.0 — upgraded from 1.19
  • Fleet cleanup — 33.2GB freed across 6 machines
  • SWAP optimized — UA 2114→320GB, AM4 837→320GB (~2.3TB NVMe reclaimed)
  • think=false routing — 21 LiteLLM routes patched

Previous: v6 — 🇹🇭 TURBO-Fable Release (Oct 25, 2026)

Current production model: TURBO-Fable DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU

Spec Value
Parameters 27.78B (27,781,081,984)
Layers 1,204
Format MLX nvfp4
Size 21.58 GB
GGUF Q4_K_M 18.50 GB
Context 8,192 tokens
Max Predict 4,096 tokens

Benchmark Results (Verified)

Thai Language Gate — 20/20 = 1.00 ✅

  • 10 questions, all scored 2/2
  • Thai character range: 292–1,064 per answer
  • Han (Chinese) leak: 0 on every question

Han Leak Test — 0/5 = 0.00 ✅

  • 5 dedicated leak probes
  • Zero Chinese characters in any response

Street Test — 3/3 PASS ✅

  • Thai chars > 50 per response
  • Han chars < 3 per response
  • Byte length > 200 per response

Inference Speed

  • Average: 18.5 tok/s (range 17–20)
  • Hardware: Apple M4 Pro (Mac Mini), MLX nvfp4

Recommended Parameters (GOD CONFIG v10.0 locked_params)

temperature: 0.6
top_p: 0.95
top_k: 20
repeat_penalty: 1.0
presence_penalty: 0.0
num_ctx: 8192
num_predict: 4096

Full bench data: benchmarks/turbo_fable_bench.yaml · Config: configs/config_g_v7.0.yaml

Promo art (ComfyUI · flux1-dev-fp8 · 1024×1024): 1 · 2 · 3

What's New in v5

🇹🇭 Fable 732 production swap (Oct 10, 2026)

  • Fable 732 (DavidAU Fable-735-882, nvfp4 MLX, digest 5bd9e2dc6f01) replaces Fable 709-L under the same qwen38-uni tag
  • 40-prompt bench: Thai 20.8 tok/s · English 21.5 tok/s · Han 0/40 direct (LiteLLM path: 1/40 leaked, fix tracked)
  • v5 benchmark card · Han detail · bench JSON

What's New in v4

🧠 Coding Stack (NEW)

  • GLM-5.3 (cloud) — Primary coding model: 71/72 benchmark, 8/8 tasks runnable
  • MiMo-V2.6 Q4 (local 9B) — Draft accelerator: 37 tok/s, zero network latency
  • Verified benchmark: 8 real coding tasks (Python, MQL5, WordPress, Bash, Node.js)
  • Rule: MiMo output never ships without test verification or GLM/human review
  • Full benchmark report →

⚡ Turbo Inference

  • Fable 732 (DavidAU Fable-735-882, Qwen3.8-27B base) — nvfp4 MLX
  • MTP speculation — 2-3× effective throughput, acceptance 0.63-0.79
  • Thai quality: 2.00 — zero Han leak, completion-gated + stacked-mark
  • think=true mode available (clean Thai reasoning)

🎙️ Production Channels

  • 5 channels + LINE bot — web chat, n8n content pipelines, live-stream Q/A overlay, chart-vision overlay, LINE (cloud fallback disclosed)
  • Live stream overlay — Restream chat → vision analysis → 4-line Thai response → Browser Source
  • Chart analysis — CDP capture → 732 vision → real-time commentary

🎵 Creative Tools

  • Image: ComfyUI + FLUX Dev fp8 (3 GPU workers via comfy-router)
  • Voice: SiangTTS (VoxCPM2 LoRA, reference-only mode)
  • Music: YuE2-3B instrumental + ACE-Step 1.5 Thai vocal

GOD CONFIG v11.0

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.

Full per-model compliance table and protected deviations: GOD_CONFIG.md

Architecture

Architecture

flowchart TB
    subgraph USERS["Channels (4 BOM + LINE hybrid)"]
        WEB["Web chat"]
        N8NP["n8n content pipelines"]
        LSO["Live-stream Q/A overlay"]
        CVO["Chart-vision overlay"]
        LINE["LINE bot (cloud fallback disclosed)"]
    end

    CF["Cloudflare Tunnel<br/>(HTTPS, key-gated)"]

    subgraph VPS["VPS — Singapore (Debian 12, Docker)"]
        OW["Open WebUI v0.13.2"]
        LL["LiteLLM proxy 1.108.3<br/>(routing, virtual keys, budgets, fallback)"]
        PG[("Postgres 16")]
        QD[("Qdrant 1.20.0")]
        VY[("Valkey 8")]
        SX["SearXNG (localhost only)"]
    end

    TS{"Tailscale mesh VPN"}

    subgraph MAC["Mac Mini M4 Pro 48GB — 'the brain'"]
        OL["Ollama 0.40.14 (MLX backend)"]
        M1["Fable 732 (27B nvfp4)<br/>~21 tok/s · MTP 2-3×"]
        MIMO["MiMo-V2.6 Q4 (9B)<br/>37 tok/s · coding drafts"]
        EMB["bge-m3 embeddings"]
    end

    subgraph I9["i9 Windows 128GB · 2× RTX 5060 Ti"]
        CU0["ComfyUI worker — GPU0 :8188"]
        CU1["ComfyUI worker — GPU1 :8190<br/>(PuLID face pin via /generate_pulid)"]
        TTS["VoxCPM2 (Thai TTS)"]
        RR["Reranker (qwen3-reranker-4b)"]
    end

    subgraph AMD["AMD render box · RTX 5060 Ti"]
        CUA["ComfyUI worker — 3rd oven"]
        ZCODE["zcode (GLM-5.3 coding)"]
    end

    subgraph NAS["Synology DS725+"]
        CR["comfy-router (FastAPI :8788)<br/>image render router"]
        MON["Uptime Kuma + hub joblog"]
        BAK["Backups + Vaultwarden"]
    end

    LINE & WEB & LSO & CVO --> CF
    N8NP --> CR
    CF --> OW & LL
    OW --> LL
    LL --> PG & VY
    OW --> SX
    OW --> QD
    LL <-->|"OpenAI-compatible, /v1"| TS
    CR <-->|"render API :8788"| TS
    TS --> OL
    TS --> CU0
    OL --> M1 & EMB
    OW -->|"image gen · images.py"| CR
    CR -->|"least-busy worker"| CU0 & CU1 & CUA
    OW -->|TTS| TTS
    OW -->|RAG rerank| RR
Loading

Coding Benchmark

MiMo-V2.6 Q4 (9B local) vs GLM-5.3 (cloud) — 8 identical tasks, no system prompt

Model Score Runnable Avg Time
MiMo-V2.6 Q4 (local) 49/72 2/8 66.3s
GLM-5.3 (cloud) 71/72 8/8 N/A

MiMo safe zone: Simple Python drafts with test assertions, scaffolding under review Keep on GLM: MQL5/EA, WordPress/PHP, Bash, timing-critical code, review-free deliverables

Detailed benchmark results →

Cost

Item Monthly Daily
Mac Mini M4 Pro electricity ~$16 $0.54
VPS (Singapore) $12 $0.40
Cloudflare (free tier) $0 $0
Tailscale (free tier) $0 $0
Total self-hosted ~$28 ~$0.94
Cloud API equivalent ~$1,500 ~$50

Savings: ~98% vs cloud API

Quick Start

git clone https://github.com/siamcafe/bomllm.git
cd bomllm
./scripts/bootstrap.sh

Fleet

Machine Role Key Specs
Mac Mini M4 Pro Brain (Primary Inference) 48GB, Ollama 0.40.16 fleet daemon, TURBO-Fable + MiMo Q4, MLX nvfp4; M5 Ultra 512GB on order
C2 VPS Singapore Gateway (Frontend) Docker, OWUI v0.13.3, LiteLLM 1.108.8 LTS, Valkey 9.1.3, SearXNG
UA Ubuntu Primary Inference + Staging 123Gi, RTX 5080 16GB + RTX 5060 Ti 16GB, Ollama 0.40.14, gpt-oss-120b, LiteLLM 1.108.8
AM4 Ubuntu ComfyUI Primary + Hub 121Gi, RTX 5080 + iGPU, Ollama 0.40.13, ComfyUI v0.45.0 x2, LiteLLM 1.108.8, hub services
i9 Windows GPU Worker 128GB, 2x RTX 5060 Ti, ComfyUI v0.45.0 (8188/8189/8190), TTS, Music, EA
AM5 Windows Coding + Render zcode GLM-5.3, ComfyUI 3rd worker
NAS DS725+ Infra comfy-router, monitoring, n8n, backups
CAT Server Web Hosting DirectAdmin, 5 WordPress sites, MySQL

Fleet Status (Nov 14, 2026)

Machine Role Ollama Status
Mac Mini M4 Pro Production qwen38-uni 0.40.16 ✅ LIVE
Ubuntu AI (UA) gpt-oss-120b + ComfyUI 0.40.14 ✅
AM4 QC Stack + Hub 0.40.13 ✅
C2 (Contabo) OWUI + LiteLLM — ✅
i9 ComfyUI + MT5 + llama.cpp — ✅
AM5 Command station 0.40.13 ✅
NAS Hub relay + n8n — ✅
CAT Web hosting (DirectAdmin) — ✅

Production pipeline: OWUI v0.13.3 → LiteLLM v1.108.8 LTS → llm-smart → gpt-oss-120b

QC System (AM4)

Deployed Oct 31 — 10/10 PASS:

  • QC Ollama :11440 (GPU0 isolated) — 4 models
  • QC Whisper :18091 (typhoon-whisper-turbo)
  • QC ASR :18092 (typhoon-asr-realtime, CPU)
  • QC Gateway :18090 (LAN accessible)

Pinned Versions

  • OWUI: v0.13.3
  • LiteLLM: v1.108.8 LTS (⛔ v1.107.0 COMPROMISED; 1.109.x forbidden line)
  • Valkey: 9.1.3
  • Ollama: 0.40.16 (Mac fleet daemon), 0.40.15 (i9), 0.40.14 (UA), 0.40.13 (AM4) — unifying on 0.40.16
  • Qdrant: v1.20.0
  • ComfyUI: v0.45.0 (fleet-wide, incl. i9:8189)
  • Cloudflared (Mac): v2026.10.7
  • BOM AI: v2.1.8 (Pro = GGUF Q4_K_M)
  • Gateway: rc3 (stream resume 0/272)
  • keep_alive: "24h" (NEVER "-1")
  • IRON Rules: 57+

License

MIT — clone, run, deploy. See LICENSE.

Contributing

See CONTRIBUTING.md.



Keywords

Thai LLM self-hosted AI Mac Mini M4 Pro BOM AI v2.1.8 GGUF Q4_K_M gateway rc3 local inference TURBO-Fable BOM LLM Thai NLP fleet management ComfyUI FLUX open-source AI MLX Ollama LiteLLM Open WebUI GOD CONFIG production AI self-hosted ChatGPT alternative Thai language model low-cost AI inference IRON rules AI operations fleet automation

🇹🇭 ภาษาไทย

BOM LLM v18.0.0 — Fleet Sync & Documentation Release

แทนที่ค่า API คลาวด์ $50/วัน ด้วย Mac Mini บนโต๊ะ $0.54/วัน (ค่าไฟจริง)

v18.0.0 ใหม่:

  • ซิงก์ GOD CONFIG v12 + GOD BLUEPRINT v8 เข้า repo — keep_alive "24h" ทุกเส้นทาง (ห้าม -1)
  • สแกนลึก Grok 8 เครื่อง / 462 ไฟล์ แก้ 8 จุด — ลบ typhoon25-content ถาวร
  • LiteLLM 1.108.8 LTS + ComfyUI v0.45.0 + Ollama 0.40.16
  • BOM AI v2.1.8 — ระดับ Pro เป็น GGUF Q4_K_M ทุกแพลตฟอร์ม (เดิม MLX เฉพาะ Mac)
  • Gateway rc3 แก้ตัดคำตัวอักษรจีน 0/272 + ปิดค้นหาเว็บเป็นค่าเริ่มต้น
  • Fleet: RAM 1,200+ GB · GPU 8 · เครื่อง 8 · อุบัติเหตุ 34 · IRON Rules 57+

ตอนนี้ทำได้ครบ: ตอบแชท, เขียนโค้ด, สร้างรูป, สร้างเสียง, สร้างเพลง, ถ่ายทอดสด — ทั้งหมดรันเครื่องตัวเอง

v17.0.0 ใหม่:

  • Open WebUI v0.13.2 + LiteLLM v1.108.3 LTS + Valkey 9.1.3 (แก้ช่องโหว่ security และ error 500)
  • ย้ายเซิร์ฟเวอร์ Contabo C2 — disk latency จาก 133ms เหลือ 4ms (เร็วขึ้น 90%)
  • bom_price_pipe ปรับเป็น async ไม่บล็อก event loop + bom-writer ล็อก num_ctx 16384
  • ตรวจสอบ fleet เชิงลึก 8 เครื่อง — 25 findings แก้จุดวิกฤต 3 จุด
  • GOD CONFIG v11.0 คงเดิม + แบนเนอร์ใหม่ธีมน้ำเงินเข้ม-ทอง

v16.0.0 ก่อนหน้า:

  • สำรวจระดับ atom ครบ 8 เครื่อง (225 ไฟล์ / 54,600 บรรทัด) + ออกแบบ 23 optimization actions
  • GOD CONFIG v11.0 + GOD BLUEPRINT v7.0
  • Price Router v2 + แก้ live-router เป็น qwen38-uni
  • LiteLLM ทั้ง fleet ล็อกเป็น v1.106.5 (v1.107.0 ยังแบน)
  • IRON Rules 44 + แบนเนอร์พรีเมียมธีมดำ-ทองใหม่

v13.0.0 ก่อนหน้า:

  • ปรับปรุง fleet ทั้ง 8 เครื่องตาม K3 blueprint (35 actions)
  • ราคาทอง XAUUSD เรียลไทม์จาก MT5 API — แก้ hallucination ราคาเก่าถาวร
  • GOD CONFIG v10.0 + GOD BLUEPRINT v6.0
  • ล็อกการเข้าถึง Netdata/Exporter ให้แคบลง ปลอดภัยขึ้น
  • Ollama 0.40.14 (Mac เรียบร้อย, AM4/i9 รอหน้าต่าง)

v12.0.0 ก่อนหน้า:

  • ตรวจสอบ fleet ครบทั้ง 8 เครื่อง (Grok 4.7 full audit)
  • ComfyUI ทั้ง fleet เป็น v0.43.0 เหมือนกันหมด (7 instances)
  • Kuma monitoring v2.5.7 บน AM4 + NAS
  • แก้ chat residuals — keepalive timeout + web-search config
  • LiteLLM incident 30 — กู้คืน models ครบ 28 ตัว
  • เวอร์ชันหลักคงเดิม: OWUI v0.13.0, LiteLLM v1.106.5, GOD CONFIG v9.0

v11.0.0 ใหม่:

  • Open WebUI v0.13.0 — อัปเดตความปลอดภัย ระบบ mods 18 ตัวใช้งานครบ
  • แก้ประสิทธิภาพเซิร์ฟเวอร์ C2 — SSH guard ถาวร ตรวจ LiteLLM DB จัดการ pollers (TTFB 0.093 วินาที)
  • GOD BLUEPRINT v5.2 + HOW-TO v4.2 — คู่มือใหม่ §3.13-§3.17
  • GOD CONFIG v9.0 ไม่เปลี่ยนแปลง (พารามิเตอร์ 7 ค่าเดิม)

v10.0.0 ก่อนหน้า:

  • OWUI v0.12.5 + Theme Designer Pro 1.19.0 — 12/12 QC PASS
  • Fleet cleanup — คืนพื้นที่ 210.43 GB ทั่ว fleet
  • Docker cleanup — ลบ 15 images ไม่ใช้ (19.33 GB)

v7 ก่อนหน้า:

  • GOD CONFIG v5.5 — 9 ค่าปรับ production
  • Open WebUI v0.12.4
  • LiteLLM v1.106.5 — patch security (v1.107.0 supply-chain compromised)
  • Ollama 0.40.10

Self-hosted AI ที่ใช้งานจริงทุกวัน ไม่ใช่ของเล่น 💪

#BOMLLM #SelfHostedAI #ThaiAI

About

Self-hosted Thai AI infrastructure — Mac Mini M4 Pro running TURBO-Fable 27B at 18.5 tok/s for $0.54/day. Fleet of 9 machines, GOD CONFIG v7.0, 40 IRON safety rules. Production-grade Open WebUI + LiteLLM + Ollama MLX stack.

Topics

Resources

Contributing

Security policy

Stars

48 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages