`moonshotai/kimi-k3` listed in /v1/models but unroutable: "not found in any configured provider"

[PhilipSkeete]

GET /v1/models on cloud-api.near.ai returns moonshotai/kimi-k3 (is_ready: true, fp4, $3.30/$16.50 per Mtok, 1M context). A minimal completions call against it fails:

Request:

POST https://cloud-api.near.ai/v1/chat/completions
{"model":"moonshotai/kimi-k3","messages":[{"role":"user","content":"Say OK"}],"max_tokens":10}

Response:

{"error":{"message":"Model 'moonshotai/kimi-k3' not found in any configured provider","type":"invalid_request_error","param":"model","code":null}}

Control: the identical request with moonshotai/kimi-k2.6 returns 200 with a normal completion — so the API key, request shape, and Moonshot lane are all fine; kimi-k3 specifically has no provider wired despite the catalog advertising it as ready.

Observed 2026-08-27 ~08:00 PT.

If this is a deploy-in-progress (noting today is your weekly release day), feel free to close once routing lands — filing since the catalog already serves the entry as is_ready: true.

[PhilipSkeete]

Confirming this is resolved: as of 2026-08-31 (~11:35 PT), the identical request returns 200 with a normal completion, and a full probe suite (streaming, tool-shaped prompts, vision, long prefill) ran clean — no routing errors. Fine to close from our side.

One observation from the verification run, for the record rather than as a new ask: the lane serves with a flat ~7.4s median TTFT that is payload-independent (same floor at ping-size and 3k-token prompts; 14s at ~50k prefill), decode ~26 tok/s, and reasoning-heavy turns at ~100-120s total. For comparison, the same model on another provider we benchmark against (Provider-B) measures ~1.1s TTFT / 66 tok/s on the identical suite. The flat floor looks like per-request overhead on the attested-3p hop rather than model compute. We also noticed /v1/models/{id}/providers returns not_implemented, so there's currently no way to see which attested third party serves a lane like this — that visibility would fit the verifiability story and would have shortened our diagnosis here.

Raw medians (n=4 per probe, 2026-08-31): ping 7.4s ttft; cold 7.3s; warm 7.4s; gen 7.7s ttft / 26 tok/s; ~50k prefill 14.2s ttft; vision 10.1s ttft. Happy to share the JSONL if useful.