feat: add nvidia/moonshotai/kimi-k3

[guillaumegay13]

✨ What changed

• Adds moonshotai/kimi-k3 (nvidia) to the catalog, params cloned from nvidia/deepseek-v4-flash-0731.yaml.
• Hosted serving: the model is developed by moonshotai (OpenRouter: moonshotai/kimi-k3); this entry documents nvidia's serving of it, probed against nvidia's endpoint.

💭 Why

The provider serves it but the catalog doesn't list it. Param surfaces differ per serving provider, so each host gets its own probed entry (models.dev-style).

📝 Notes

• Liveness ✅ only: the model answers on its real endpoint; unverified params are listed in the evidence table below.
• Draft PR, auto-proposed by the modelparams agent.

Param verification evidence

param	check	ok	detail
(baseline)	accepted	✅	1-shot answers 200
temperature	min=0	✅
temperature	max=1	✅
temperature	range policed	✅	temperature=2 rejected
top_p	max=1	❌	Validation: top_p is immutable for this model and must be 0.95, got 1
top_p	range policed	✅	top_p=2 rejected
max_tokens	min=1	✅
max_tokens	max=16384	➖	too large for a non-streaming probe — upper bound from docs
max_tokens	range policed	➖	vendor accepts 100M (caps silently) — declared max 16384 is a docs value, not API-policed
frequency_penalty	min=-2	❌	Validation: frequency_penalty is immutable for this model and must be 0, got -2
frequency_penalty	max=2	➖	{"status":429,"title":"Too Many Requests"}
presence_penalty	min=-2	➖	{"status":429,"title":"Too Many Requests"}
presence_penalty	max=2	➖	{"status":429,"title":"Too Many Requests"}
seed	min=0	➖	{"status":429,"title":"Too Many Requests"}
seed	max=18446744073709552000	➖	{"status":429,"title":"Too Many Requests"}
stop	declared	➖	type string has no probe — value carried from sibling/docs, not verified

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
modelparams.dev	Ready	Preview	Sep 7, 2026 7:21pm UTC