from typing import Literal from pydantic_settings import BaseSettings class Settings(BaseSettings): # ── Model stack — sourced from MADdegens/Models HF bucket ──────────────── # GGUFs pulled via `hf buckets cp` → `ollama create` → `ollama push` (cloud) # Roles and their Ollama tags: # # reasoner nicholasjmcleod/kimi-linear:q6_k Kimi-Linear-48B (1M ctx preprocessor) /v1/instruct /v1/chat # tandem nicholasjmcleod/harmonic-hermes:q6_k Harmonic-Hermes-9B (refine+monitor) /v1/tandem stage-2 + /v1/monitor # coder nicholasjmcleod/qwen3-coder-53b:q6_k Qwen3-Coder-53B TOTAL-RECALL (primary) /v1/code /v1/tandem stage-3 # coder_dedicated nicholasjmcleod/qwen3-coder-53b:q6_k Qwen3-Coder-53B TOTAL-RECALL (dedicated) /v1/coder # kimi_distilled nicholasjmcleod/kimi-distilled:iq4_xs Qwen3.6-35B K2.6-Distilled (reasoning) /v1/reason # genstruct nicholasjmcleod/genstruct:q6_k Genstruct-7B (structure middleware) /v1/generate + pre/post all # infinity_parser nicholasjmcleod/infinity-parser:latest Infinity-Parser2-Pro (parse midw.) /v1/parse + pre/post all # monitor_pair harmonic-hermes + qwen3-coder-53b Hermes watches / Coder remediates /v1/monitor/analyze reasoner_model: str = "nicholasjmcleod/kimi-linear:q6_k" tandem_model: str = "nicholasjmcleod/harmonic-hermes:q6_k" # coder and coder_dedicated both point to Qwen3-Coder-53B TOTAL-RECALL coder_model: str = "nicholasjmcleod/qwen3-coder-53b:q6_k" coder_dedicated_model: str = "nicholasjmcleod/qwen3-coder-53b:q6_k" kimi_distilled_model: str = "nicholasjmcleod/kimi-distilled:iq4_xs" genstruct_model: str = "nicholasjmcleod/genstruct:q6_k" infinity_parser_model: str = "nicholasjmcleod/infinity-parser:latest" # ── Monitor pair — Harmonic-Hermes (analysis) + Qwen3-Coder-53B (remediation) ── monitor_model: str = "nicholasjmcleod/harmonic-hermes:q6_k" monitor_coder_model: str = "nicholasjmcleod/qwen3-coder-53b:q6_k" monitor_poll_interval_secs: int = 30 monitor_targets: str = "http://multimodal:8000,http://coder:8001,http://research:8002,http://mcp-bridge:8004,http://ollama:11434" # ── Ollama endpoint ─────────────────────────────────────────────────────── ollama_base_url: str = "http://ollama:11434" ollama_api_url: str = "http://ollama:11434/api" ollama_api_key: str = "" openwebui_url: str = "http://localhost:3000" # ── Ollama Cloud auth ───────────────────────────────────────────────────── ollama_cloud_enabled: bool = True ollama_cloud_account: str = "nicholasjmcleod@gmail.com" ollama_cloud_device: str = "LN-8RDGA90Ultra" # ── [2] Moonshot API — primary Kimi K2.6 endpoint (free credits) ───────── moonshot_api_key: str = "" moonshot_api_url: str = "https://api.moonshot.ai/v1" moonshot_model: str = "kimi-k2-0711-preview" # ── [3] OpenRouter — K2.6 fallback ─────────────────────────────────────── openrouter_api_key: str = "" openrouter_api_url: str = "https://openrouter.ai/api/v1" openrouter_model: str = "moonshotai/kimi-k2.6" # ── [4] HuggingFace Router — tertiary compute ───────────────────────────── hf_token: str = "" hf_token_2: str = "" hf_token_3: str = "" hf_router_url: str = "https://router.huggingface.co/v1" hf_reasoner_model: str = "moonshotai/Kimi-K2-Instruct-0905" hf_tandem_model: str = "NousResearch/Hermes-3-Llama-3.1-8B" hf_coder_model: str = "Qwen/Qwen3-Coder-480B-A35B-Instruct" hf_frontier_reasoner: str = "moonshotai/Kimi-K2-Instruct-0905" hf_frontier_coder: str = "Qwen/Qwen3-Coder-480B-A35B-Instruct" # ── [5] RunPod Serverless — quaternary GPU burst ────────────────────────── runpod_api_key: str = "" runpod_s3_key: str = "" runpod_reasoner_endpoint_id: str = "kukl55t0053lob" runpod_coder_endpoint_id: str = "kukl55t0053lob" runpod_api_url: str = "https://api.runpod.ai/v2" runpod_gpu_class: str = "AMPERE_16" # ── HF Bucket — model source ────────────────────────────────────────────── hf_bucket_repo: str = "MADdegens/Models" hf_bucket_base: str = "https://huggingface.co/MADdegens/Models/resolve/main" model_cache_dir: str = "/models" # ── Search + Research tools ─────────────────────────────────────────────── perplexity_api_key: str = "" perplexity_default_model: str = "sonar-pro" exa_api_key: str = "" tavily_api_key: str = "" smithery_api_key: str = "" # ── Compute routing ─────────────────────────────────────────────────────── # cloud_priority = Ollama Cloud → Moonshot → OpenRouter → HuggingFace → RunPod compute_strategy: Literal[ "cloud_priority", "round_robin", "load_based", "local_first", "hf_first", "runpod_first", "openrouter_first", ] = "cloud_priority" # Weights calibrated to actual subscription capacity (May 2026): # Ollama Cloud Pro ($20/mo): 3 concurrent sessions, ~$0.67/day GPU time # Moonshot Tier 0 (free): 1.5M tokens/day FREE = ~750 K2.6 calls/day # Moonshot Tier 1 ($10 rc): 50 concurrent / 200 RPM after recharge # OpenRouter (PAYG): no platform limits; big coding + overflow # HF Router ($49/mo): $2 Personal + ~$4-6 Teams = ~$8/mo credits # + 50 ZeroGPU Spaces (Teams org) # RunPod Community ($25 dep):~73 hrs RTX 4090 total, scale-to-zero local_weight: int = 38 moonshot_weight: int = 37 openrouter_weight: int = 15 hf_weight: int = 8 runpod_weight: int = 2 # ── Shared infrastructure ───────────────────────────────────────────────── redis_url: str = "" database_url: str = "" # ── Server ──────────────────────────────────────────────────────────────── port: int = 8000 log_level: str = "info" local_max_queue: int = 8 hf_max_queue: int = 8 request_timeout_secs: int = 300 class Config: env_file = ".env" env_file_encoding = "utf-8" # ── Helpers ─────────────────────────────────────────────────────────────── def active_hf_token(self) -> str: return self.hf_token or self.hf_token_2 or self.hf_token_3 or "" def has_ollama(self) -> bool: return True def has_moonshot(self) -> bool: return bool(self.moonshot_api_key) def has_openrouter(self) -> bool: return bool(self.openrouter_api_key) def has_hf(self) -> bool: return bool(self.active_hf_token()) def has_runpod(self) -> bool: return bool(self.runpod_api_key and self.runpod_reasoner_endpoint_id) def has_perplexity(self) -> bool: return bool(self.perplexity_api_key) def has_exa(self) -> bool: return bool(self.exa_api_key) def has_tavily(self) -> bool: return bool(self.tavily_api_key) def has_monitor(self) -> bool: return True def all_local_models(self) -> list[str]: return [ self.reasoner_model, self.tandem_model, self.coder_dedicated_model, self.kimi_distilled_model, self.genstruct_model, self.infinity_parser_model, ] settings = Settings()