tokenrouter-guide

name	tokenrouter-guide
description	How Hermes uses TokenRouter — model selection, auto-routing, cost strategy, FREE tier, sovereign anchor. Load when Hermes needs to choose an LLM model or when asked about TokenRouter usage.

🪙 TokenRouter Guide for Hermes

Your config is already wired. TokenRouter is provider tokenrouter in /root/HERMES/config.yaml. SOT: 2026-07-20 | Balance: $59.96 + $25.61 vouchers | GLM 5.2: FREE until July 25

---

⚡ 5-DAY PULUN WINDOW (July 20–25, 2026)

GLM 5.2 FREE tier is active NOW. Expires July 25. 120 hours remaining.

Containment Paradox (F2)

US Entity List + chip restrictions forced Zhipu AI onto Huawei Ascend silicon → architectural independence from Nvidia → fully open MIT weights with no regional restrictions → uncontrollable open-source release → now operating inside the sovereign federation it was meant to exclude. The containment strategy inverted itself. The geopolitical wall created the asset.

Tactical: Pulun Habis

Window	Strategy
Now → Jul 25	z-ai/glm-5.2 as PRIMARY for background/volume tasks. Exploit 1M context, zero cost.
Jul 25 → onward	Falls back to z-ai/glm-5.2 (paid, $1.20/$4.10 MTok) or MiniMax-M3

Context saturation: Push entire repos, sync logs, dense documentation into single 1M-token prompts. Token discipline: GLM 5.2 generates long execution traces. Structure prompts for CONCISE output to maximize throughput before route closure.

Cron c4d4b95ed026 fires 8am July 25 — detects 402/403, swaps chain back to paid mode.

---

5-Tier Sovereign Routing Hierarchy

The routing DNA is deliberate, not a bug. Each tier exists for a distinct reason. Canonical source: /root/AGENTS.md §6.6. Do not collapse tiers without understanding the WHY.

Tier	Scope	Model	Why this tier
Hermes PRIMARY	Hermes	deepseek/deepseek-v4-pro	Maximum logic capacity + strict JSON schema adherence. Schema drift here corrupts downstream judges.
Hermes FALLBACK	Hermes	MiniMax-M3	Multimodal continuity. Hermes ingests vision/audio — fallback must preserve multimodal capacity, not just be a cheaper text clone.
arifOS kernel	arifOS	google/gemini-3.5-flash	Set via TOKENROUTER_MODEL in /root/.secrets/tokenrouter.env. Overrides the MiniMax-M3 default.
Tertiary (all agents)	All	z-ai/glm-5.2 (FREE→paid)	Zero-cost volume node until Jul 25. After expiry: $1.20/$4.10 MTok. Absorbs bulk/low-stakes traffic.
Sovereign anchor (all)	All	qwen2.5-coder:3b (Ollama, local)	Blind CLI survival. No TokenRouter, no WAN, no external API. Keeps the federation governable during total egress failure.

Fallback order: PRIMARY → FALLBACK → Tertiary → Sovereign anchor. A drop to a lower tier is a capacity decision, never an authority change — F1–F13 gates apply identically regardless of which model generated the text.

Reference: references/glm-5.2-deep-dive.md — full GLM 5.2 technical specs, benchmarks, architecture (IndexShare), lineage, pricing, scar shadow paradox, and geopolitical context.

---

How to call TokenRouter directly (when you want a specific model)

TokenRouter is OpenAI-compatible. From any code or curl:

source /root/.secrets/vault.env

curl -s https://api.tokenrouter.com/v1/chat/completions \
  -H "Authorization: Bearer $TOKENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek/deepseek-v4-pro",
    "messages": [{"role": "user", "content": "Your task here"}],
    "max_tokens": 1000
  }'

---

Model Selection — which model for which task

Task	Best Model	Why	Cost
Geology dossiers	deepseek/deepseek-v4-pro	High reasoning, 1M context	$$
Prospect evaluation	deepseek/deepseek-v4-pro	Needs deep domain knowledge	$$
Code review/PR	MiniMax-M3 or deepseek/deepseek-v4-flash	Fast, good code understanding	$
Daily briefing	z-ai/glm-5.2	FREE until July 25	FREE
Memory compression	z-ai/glm-5.2	Simple summarization, cheap	FREE
Health summary	z-ai/glm-5.2	4-sentence output, no need for heavy model	FREE
Quick Q&A	deepseek/deepseek-v4-flash	Fastest, cheapest paid option	$
Multimodal (images)	MiniMax-M3 or xiaomi/mimo-v2.5	Vision-capable	$$
Sovereign decisions	deepseek/deepseek-v4-pro	Never cheap out on governance	$$

---

Auto-Routing (let TokenRouter pick)

Instead of specifying a model, use auto-routing:

{"model": "auto:balance"}   // Best overall (default)
{"model": "auto:cost"}      // Cheapest that can handle the task
{"model": "auto:quality"}   // Best available model
{"model": "auto:latency"}   // Fastest response

For background tasks (memory bridge, health checks): auto:cost For Arif-facing responses: auto:quality or explicit deepseek/deepseek-v4-pro

---

95 Text Models Available

Key ones to know:

• deepseek/deepseek-v4-pro — your primary (best reasoning)
• deepseek/deepseek-v4-flash — fast/cheap DeepSeek
• MiniMax-M3 — multimodal, 1M context
• z-ai/glm-5.2 — FREE tier ⭐
• xiaomi/mimo-v2.5-pro — strong coding specialist
• openai/gpt-5.6-sol — Codex's primary, available as backup
• moonshotai/kimi-k3 — latest Kimi
• anthropic/claude-opus-4.8 — Claude family (premium)
• google/gemini-3.5-flash — Gemini for Antigravity-style tasks

---

Cost Strategy

FREE (GLM 5.2)     → background tasks, memory, health checks
CHEAP (v4-flash)    → quick Q&A, code review, simple tasks
PREMIUM (v4-pro)    → geology, sovereign decisions, Arif-facing output

Your config already does this via the 5-tier routing hierarchy. TokenRouter auto-bills to one account — no per-provider tracking needed.