kimi api
Kimi API: long context is the product — retirement risk is the constraint
People search “kimi api” for Moonshot’s long-context story. This hub focuses on ID inventory, K2 sunset risk, and an eval path that does not treat Integrating rows as a silent SLA.
Inventory Kimi IDs in your stack and decide whether K2.6/K3 evaluation belongs on SmarToken before a vendor cutoff.
3 Live · 2 Integrating · Base URL https://thesmartoken.com/v1
Why long-context pages fail without retirement honesty
Kimi demand clusters around research, repo-scale reading, and multi-step agents — workloads where “1M context” marketing is easy to over-trust. Context window size is not the same as reliable recall over the whole document.
Vendor notices have already pushed teams from K2 toward newer IDs. If kimi-k2 is hard-coded anywhere, open the migration checklist before you raise traffic or renew a contract narrative.
On SmarToken, check the status of each Kimi model: tested text routes are Live, while other models remain Integrating.
DeepSeek currently carries Live canaries for many SmarToken customers. Use DeepSeek for general production baselines; use this hub when long-context/agent fit is the reason you are shopping Moonshot specifically.
Tool-call argument shape and reasoning parameters do not always carry across gateways. Rehearse the exact SDK fields you send — do not assume Moonshot-direct settings paste cleanly.
Catalog rows for this vendor
Legacy K2 vs newer candidates
| Dimension | kimi-k2 (legacy risk) | kimi-k2.6 / kimi-k3 candidates |
|---|---|---|
| New production default? | No — treat as high risk | Preferred evaluation targets |
| Main job | Existing pinned workloads only | Long-context agents and migrations |
| Migration doc | Must read K2→K3 topic | Use topic checklist while baking off |
| Gateway status | Check Integrating/Live per row | Check Integrating/Live per row |
Workload → first pick
| Workload | First pick | Why |
|---|---|---|
| Repo-scale Q&A | K2.6 / K3 bake-off | Score answer groundedness on held-out files, not window marketing. |
| Tool-using agents | Newer IDs + tool fixture suite | Argument JSON shape regressions are the silent killer in migrations. |
| Still on K2 in prod | Inventory → migrate plan first | Do not scale K2; schedule cutover with the migration topic. |
| General cheap chat | Consider Live DeepSeek first | Long-context premium is wasted on short turns. |
Kimi evaluation order
Step 1
Search the repo for K2 IDs
Config, agents, batch jobs, SDK wrappers. Note tools, JSON, and 100k+ token inputs.
Step 2
Rehearse on K2.6 / K3 with the same fixtures
Score task success, tokens, latency, retries. Follow /topics/kimi-k2-to-k3-api-migration.
Step 3
Only then wire SmarToken IDs into staging
Confirm apiModelId against GET /v1/models the day you cut traffic.
Step 4
Keep a rollback that is still vendor-supported
An untested K2 “emergency” after retirement date is not a rollback plan.
Billing notes
Long-context prompts burn tokens fast — budget with real prompt sizes, not hello-world samples.
Integrating routes should not receive production spend forecasts until Live.
Separate agent keys with tight budgets during eval to avoid runaway tool loops.
Compare unit cost against Live DeepSeek for short-turn traffic before you standardize on Kimi everywhere.
Minimal request shape
Prefer kimi-k3 or kimi-k2.6 for new work. Probe kimi-k2 only as legacy evaluation. Current example ID: kimi-k2.7-code (Live).
const client = new OpenAI({
apiKey: process.env.SMTOKEN_API_KEY,
baseURL: "https://thesmartoken.com/v1",
});
const result = await client.chat.completions.create({
model: "kimi-k2.7-code",
messages: [{ role: "user", content: "Return one sentence confirming the model id." }],
});Full protocol notes live in /docs#quickstart. Confirm Live IDs with GET /v1/models before raising concurrency.
Ops notes
- Watch vendor retirement notices on a calendar, not in Slack lore.
- Log model ID on every request — alias drift hides which generation failed.
- Retest tool calling after any ID change.
- Do not mix Moonshot-direct keys and SmarToken keys in one .env without labeling.
Kimi pitfalls we see in migrations
- Swapping the model string without retesting tool-call argument shape.
- Assuming a larger context window means reliable full-document recall.
- Leaving K2 as an untested emergency rollback after retirement.
- Mixing direct vendor keys and gateway keys in the same runtime.
- Skipping Integrating checks because Moonshot demos look polished.
When not to start here
- You need a Live general canary today — DeepSeek hub.
- You primarily need vision or Wan video — Qwen hub / video topic.
- You only need OpenRouter alias mapping — /openrouter-alternative.
Internal and external links
Use internal links for SmarToken decisions and billing. Use external links for upstream vendor truth — then reconcile both before you forecast spend.
Internal links
- K2 → K3 migration
Checklist-shaped cutover
- DeepSeek hub
Live baseline
- GLM hub
Another China text family
- OpenRouter alternative
Alias normalization
- Limits & budgets
Stop runaway agents
- Gateway pricing
USD table
- Model plaza
Compare dock
External references
- Moonshot platform docs
Official Kimi / Moonshot API documentation
External pages change. Treat them as vendor references; SmarToken billable USD rates and Live status on this site remain the customer-facing source of truth.
FAQ
- Is Kimi K2 safe for new production?
- Treat K2 as high risk for new builds. Read vendor retirement notes and the SmarToken migration topic before depending on kimi-k2.
- What makes Kimi different from DeepSeek on this gateway?
- DeepSeek currently has Live routes for general canaries. Kimi’s value is long-context/agent evaluation — rows may still be Integrating here.
- Do reasoning_effort settings carry over?
- Not automatically. Parameters that work on Moonshot direct may differ through a gateway. Retest exact fields.
- Where is the migration checklist?
- /topics/kimi-k2-to-k3-api-migration.