Configure Kimi (Moonshot) with Hermes Agent: 128K-1M long context

Why Kimi

Kimi is developed by Moonshot AI, the first Chinese LLM vendor to market long context as a primary capability. For Hermes Agent users, the value is clear: when GLM-4 / Claude 128K-200K windows fill up, Kimi-1M takes the baton.

• 🪟 1M tokens of context: ~750K Chinese characters, enough to hold a complete medium codebase (100-150K LOC) or a 200-page PDF.
• 🔍 High long-context recall: measured Kimi-1M needle-in-haystack accuracy > 95% across the 500K-1M token range, clearly ahead of Chinese peers that nominally support 128K but degrade past 60K.
• 🇨🇳 Stable in-China access: moonshot.cn direct, Beijing / Shanghai / Shenzhen first-token latency ≈ 400-500ms. No proxy required.
• ⚡ Streaming-friendly: long-output streaming is stable (low jitter, few drops), good fit for Hermes Gateway real-time echo scenarios.

Kimi is not a GLM replacement — it’s a complement. Daily: GLM-4-Plus. Long-context (> 60K) tasks: switch to Kimi. Hermes lets you configure multiple providers in parallel.

Get a Moonshot API key

Moonshot Open Platform flows like Zhipu — about 5 minutes end to end.

1. 1. Sign up at Moonshot Open Platform

   Visit https://platform.moonshot.cn → sign up with phone → real-name verify. Business accounts get higher RPM ceilings.

2. 2. Generate and copy an API key

   After login → Account Settings → API Key Management → New API Key. Copy the key (format: sk-xxxxxx) to a password manager. The key is shown once only.

3. 3. Top up (optional)

   New accounts get a 15 RMB trial credit. Kimi-128k input is ≈ 0.060 RMB / 1K tokens, so the trial covers ~250K tokens (about one medium PDF). To continue, top up via Finance Center (10 RMB minimum).

Recommend burning trial credit on 1-2 long-context tasks before paying — see if the capability fits your workflow.

Configure Hermes to use Kimi

Hermes ships a built-in moonshot adapter using the official OpenAI-compatible endpoint. 3 commands and you are done.

hermes config set provider.moonshot.api_key "YOUR_KIMI_KEY"

1. 1. Write the API key

   Once the command runs the key lands encrypted on disk — it will not appear in git diffs. With multiple providers, Kimi and GLM keys coexist without interference.

2. 2. Pick the default model

   hermes config set provider.moonshot.model moonshot-v1-128k (recommended: best 128K cost/perf). Switch to moonshot-v1-1m only when ultra-long context is needed.

3. 3. Make Kimi the global default OR per-Skill

   Global switch: hermes config set provider moonshot. Per-Skill switch (keep GLM default, use Kimi only for long-context tasks): add provider_override: moonshot inside the Skill meta.yaml (see /skills).

Run hermes doctor --provider to verify — a ✅ within 5 seconds means the connection is good.

Model selection (8k / 32k / 128k / 1M)

Moonshot exposes 4 tiers, differentiated by context window and pricing. They all look like "Kimi" but performance and cost vary by an order of magnitude — picking wrong can multiply your bill 10×.

• moonshot-v1-8k: 8K context, input ≈ 0.0024 USD/1K tokens, output 0.0024. Best for short chats, simple classification, prototypes. Lowest cost daily-chat tier.
• moonshot-v1-32k: 32K context, input ≈ 0.0048 USD/1K tokens. Best for medium-length doc summary, few-file RAG.
• moonshot-v1-128k: 128K context, input ≈ 0.012 USD/1K tokens. Hermes Agent default long-context choice, best cost/perf. Holds a 50-80K-character project doc.
• moonshot-v1-1m (kimi-1m): 1M context, input ≈ 0.04 USD/1K tokens. Dedicated model — slower and pricier but most capable. For ingesting an entire codebase or thick contract. Requires manual enable via console "Service activation".
• kimi-k1.5 (reasoning-enhanced): time-limited availability for complex reasoning tasks, pay-as-you-go.

Default to moonshot-v1-128k. Only escalate to 1M when a single call demonstrably needs > 80K. Refer to the official Moonshot pricing page for the latest → https://platform.moonshot.cn/docs/pricing .

5 long-context scenarios where Kimi shines

Once Kimi is wired in, these 5 task types are where it actually outperforms shorter-context options. Each maps to a built-in / recommended Hermes Skill.

• Scenario 1 — Whole-codebase audit: Skill code-audit + Kimi-1M. Drop ~100K LOC into a single context window so the agent can find architectural smells, circular deps, security gaps. Models that degrade after 60K (most 128K-nominal Chinese LLMs) wobble here.
• Scenario 2 — Long contract / legal doc summary: Skill doc-summarize + Kimi-128k. Feed in a 200-page PDF (~300K chars), get per-chapter summaries + key-clause risk flags. More accurate than naive chunk-and-merge RAG.
• Scenario 3 — Multi-turn research doc dialog: Skill research-chat + Kimi-128k. Import 10-20 papers, hold multi-turn dialog with cross-paper reasoning and consistent context. Hermes Memory module can compress those long sessions down to 5K on reload.
• Scenario 4 — Multi-file Skill chain: Hermes Gateway running "read 50 YAML configs → produce a comparison table → write a migration script" runs end-to-end without intermediate chunking on Kimi-128k.
• Scenario 5 — 200+ turn conversation without context loss: Hermes Chat mode with Kimi-128k + Hermes auto-context management beats GLM-4 / GPT-4o, which start forgetting after ~30 turns.

If your daily tasks stay under 60K tokens, you do not need Kimi — GLM-4-Plus is cheaper and faster. Pull Kimi out only when long context is the actual requirement.

Common errors

Four error buckets cover most Kimi day-to-day issues.

• context_length_exceeded

  原因 / Cause: The current dialog or single request exceeds the chosen model’s window. e.g. moonshot-v1-8k but the input is 12K tokens.

  修复 / Fix: Two options: (a) escalate: hermes config set provider.moonshot.model moonshot-v1-128k; (b) compress: hermes context summarize to compress prior turns. If you really need > 128K, switch to moonshot-v1-1m.

• Model not found

  原因 / Cause: Your account lacks access. moonshot-v1-1m is request-only by default.

  修复 / Fix: Console → Service Activation → request 1M model. Approval is usually 1-2 business days. For 128k and below, confirm real-name verification first.

• 401 Invalid API key

  原因 / Cause: Key copy error, expired, or generated under another account.

  修复 / Fix: Regenerate in the console, run hermes config set provider.moonshot.api_key "NEW_KEY". Moonshot keys start with sk- — do not strip the prefix.

• 429 Rate limit / TPM exceeded

  原因 / Cause: Moonshot throttles by TPM (tokens per minute), not just RPM. One long-context request can blow the TPM ceiling.

  修复 / Fix: Console → Account Center → Rate Limits → request a higher TPM. Short-term workaround: hermes config set llm.moonshot.concurrent 1 to serialize requests until you get the quota bump.