Where to Use Kimi K3: Every API Provider, Price & How to Choose

Kimi K3 is suddenly everywhere. In the two weeks since launch, Moonshot AI has pushed its 2.8-trillion-parameter model through the widest third-party distribution blitz the open-weight world has seen: independent trackers now list 10–14 API providers, from US inference clouds to European hosts to budget GPU marketplaces. Moonshot’s X feed (@Kimi_Moonshot) has been announcing new partners almost daily — and the message is clear: you no longer need Moonshot’s own API (or a $300K GPU node) to build on K3.

This guide maps the entire distribution network as of August 1, 2026 (updated September 25 for Amazon Bedrock, OpenRouter’s multi-provider routing and current pricing): every verified provider, what each one charges, which ones host in the US/EU, which are fast, which are cheap, the hyperscaler story (absent at launch, Bedrock since September 18), and a pick-the-right-one framework at the end.

The master table: every Kimi K3 provider compared

Provider	Model ID	Price (per 1M tokens)	Standout feature
Moonshot (official)	kimi-k3	$3 in / $0.30 cached / $15 out	Reference behavior, full feature set, $1 top-up unlock
Together AI	moonshotai/Kimi-K3	$3 / $0.30 / $15	Serverless now, dedicated capacity later
Fireworks AI	accounts/fireworks/models/kimi-k3	$3 / $0.30 / $15	US-hosted + zero data retention, fine-tuning
Baseten	moonshotai/Kimi-K3	Console/quote	Production-grade infra (8×GB300 replicas)
Modal	kimi-k3	$3 / $0.30 / $15	Claims 460 tok/s shared endpoint (provider-measured)
Nebius	kimi-k3	~$3/$15 class	European hosting (Amsterdam)
SiliconFlow	Kimi-K3	$3 / $0.30 / $15	OpenAI and Anthropic-compatible endpoints
Novita	moonshotai/kimi-k3	~$3/$15 class	Also an upstream for Vercel AI Gateway
DigitalOcean	—	~$3/$15 class	Fits existing DO stacks
Databricks	—	Enterprise	Lakehouse-integrated serving
OpenRouter	moonshotai/kimi-k3	$3 / $15	One key for hundreds of models (~20 K3 upstreams as of Sep 2026)
Vercel AI Gateway	moonshotai/kimi-k3 (+kimi-k3-fast)	$3 / $0.30 / $15	Multi-upstream failover, ZDR, $5 free credits/30 days
Cloudflare Workers AI	moonshotai/kimi-k3	In dashboard	Inference inside Workers/edge stack
CrofAI	crofai/kimi-k3	$2 in / $8 out — lowest tracked output rate	Budget endpoint
Chutes	moonshotai/Kimi-K3-TEE	Budget class	TEE (confidential compute) serving
io.net	moonshotai/Kimi-K3	Budget class	Decentralized GPU network
Venice	kimi-k3	Budget class	Privacy-focused app + API
RunPod	kimi-k3 (public endpoint)	Flat ~$15/M (no split)	Shared endpoint, no setup

Prices are a July 27–August 1 snapshot from provider pages and trackers (endpoints.run, Artificial Analysis). Rates, regions and capacity move fast — always confirm on the provider’s own page before committing. Update, September 25: OpenRouter’s cheapest K3 endpoints now undercut CrofAI on input — Sail Research at ~$1.35 in / $11.50 out and InferenceNet at ~$1.40 / $10.75 per 1M.

Why Moonshot is distributing K3 so aggressively

The speed of this rollout is the real story. Three moves explain it:

1. The subscription wall. K3 demand was so heavy at launch that Moonshot had to pause new subscriptions — its own infrastructure literally couldn’t onboard everyone. Every third-party provider is a pressure-release valve that still drives usage (and, via the Kimi K3 License, a separate commercial agreement with Moonshot once a Model-as-a-Service business passes $20M in revenue over 12 months).
2. The hyperscaler gap. At launch, AWS Bedrock, Azure AI Foundry and Google Vertex carried none of the Chinese open-weight flagships, so Moonshot made K3 one API key away on US and EU soil through Fireworks, Baseten, Together, Nebius and the gateways. That strategy worked: Microsoft Foundry now offers K3 via Fireworks, and Amazon Bedrock added it on September 18, 2026.
3. Open weights as the trust anchor. Because the weights are public, no provider can lock you in. If one host degrades or reprices, you switch model endpoints, not models. That makes developers far more willing to build on K3 — which is exactly what Moonshot needs to close the ecosystem gap with the closed frontier.

The result: for the first time, a frontier-class Chinese model is easier to access from a US laptop than some US models are.

The official route: Moonshot’s own API and app

Everything else is measured against this baseline:

• API: model="kimi-k3" at https://api.moonshot.ai/v1 (OpenAI-compatible). Full 1,048,576-token context, vision input, tool calling, structured output, automatic prompt caching, and reasoning_effort of low / high / max. K3 unlocks after a $1 top-up. Moonshot’s developer docs now live at platform.kimi.ai (platform.moonshot.ai redirects there); the API base URL and model ID are unchanged.
• Price: $3/M input, $0.30/M cached input, $15/M output — and Mooncake’s prefix caching means real coding workloads often bill mostly at the $0.30 rate. The pricing page also lists cache writes at $3/M for the default 5-minute TTL and $6/M for a 1-hour TTL.
• Products: the Kimi app, kimi.com, Kimi Work and Kimi Code CLI share a membership-credit system. Since the September 18, 2026 plan relaunch, Kimi Code requires the Plus tier or above (details). (Full walkthrough.)

Pick the official API when you want reference behavior, the newest features first, and no middleman. Pick a third party when you need a data boundary, an existing stack, or failover.

The major managed providers

Together AI — the day-0 partner. Serverless K3 at official pricing, with a clean upgrade path to dedicated and provisioned-throughput capacity. Choose Together when you want to start on shared infrastructure now and reserve GPUs later without changing endpoints.

Fireworks AI — the compliance pick. US-hosted serverless endpoint with zero data retention, on-demand dedicated GPUs, function calling, vision, and fine-tuning — the only major provider pairing K3 serving with a tuning path. If legal asks “where does our data go?”, Fireworks is the shortest answer.

Baseten — the infrastructure pick. Baseten’s day-0 write-up is the most transparent look at K3 serving anywhere: MXFP4 files exceeding 1.4TB, production replicas of 8× NVIDIA GB300 GPUs, validated through the Kimi Vendor Verifier on both vLLM and SGLang. No public per-token rate — console or quote — but if you’re comparing enterprise serving, this is the grown-up conversation.

Modal — the speed pick. Its shared endpoint advertises 460 output tokens/second via a custom DFlash speculator — versus ~35 tok/s measured on the standard path. Treat it as a provider benchmark, not a neutral one, but if interactive latency is your product, Modal is worth a test drive.

Nebius — the Europe pick. Amsterdam-based AI cloud, useful for EU data-residency requirements that rule out both Moonshot-direct and US hosts.

SiliconFlow — the compatibility pick. Both OpenAI-compatible and Anthropic-compatible endpoints, with documented setup for Claude Code, Cline and OpenCode — the easiest drop-in for Anthropic-format coding agents.

Also live: Novita (which doubles as a Vercel AI Gateway upstream), DigitalOcean (natural if your infra already lives there), and Databricks (K3 inside the lakehouse for enterprise data teams). Artificial Analysis’s provider tracker rounds out the list with Parasail, Makora and Wafer for price-shopping.

Gateways: one key, many models

OpenRouter exposes moonshotai/kimi-k3 behind its single-key billing layer. At launch K3 ran through one upstream provider there; by late September 2026 OpenRouter lists about 20 K3 endpoints from 17 providers, with prices ranging from roughly $1.35/$11.50 to $6/$22.50, so you now get real failover too. Watch for :free community routes — OpenRouter regularly provisions subsidized zero-cost endpoints after major launches.

Vercel AI Gateway is the more interesting K3 story right now. It routes moonshotai/kimi-k3 across multiple upstream providers (Moonshot, Novita, Baseten, Fireworks) with automatic fallback, publishes live latency/uptime stats, supports ZDR, and hosts Kimi K3 Fast — a lower-latency tier at roughly 50% higher per-token cost, requestable as moonshotai/kimi-k3-fast or a speed option. Unpaid teams get $5 of free credits every 30 days. If your app already speaks the Vercel AI SDK, this is the lowest-friction production route.

Cloudflare Workers AI lists K3 at the full 1M context in an OpenAI-compatible format — the obvious choice if your stack is already Workers, Queues and Cloudflare billing. Pricing shows in the dashboard.

The budget tier (with a caveat)

Below the majors sits a cluster of GPU-marketplace and community providers — CrofAI (lowest tracked output rate at $2.00/M input, $8.00/M output; on input, OpenRouter’s Sail Research at ~$1.35/M and InferenceNet at ~$1.40/M now go lower), Chutes (TEE confidential-compute serving), io.net, Venice, AtlasCloud, Phala, Inference.net, and RunPod (flat ~$15/M shared public endpoint, no input/output split).

The caveat: at this tier, verify what you’re actually getting — which quantization is served, real context length, uptime history, and data policy — before routing production traffic. For side projects and evaluation, though, these are legitimately cheap ways to burn through millions of K3 tokens.

What you won’t find (yet)

• Google Vertex: still no native K3 listing. Update, September 25, 2026: this section originally said all three hyperscalers were missing. Amazon Bedrock launched K3 on September 18, 2026 at $3/$15 on global routing, and Microsoft Foundry offers K3 served by Fireworks.
• Devin: Cognition’s agent doesn’t accept third-party model keys and doesn’t expose K3 as a selectable model.
• A laptop: distribution doesn’t change physics — serving K3 needs ~1.7TB of GPU memory. The providers above exist precisely because almost nobody can run this model themselves.

Speed and real cost: two honest numbers

Speed: the standard serving path measures around 35 output tokens/second with ~3.8s time-to-first-token at max reasoning (Artificial Analysis, Moonshot’s endpoint) — K3 is a deliberate, verbose thinker. Fast tiers (Modal’s speculator, K3 Fast on Vercel) buy latency back at a price premium. If your workload is batch agents, standard speed is fine; if it’s interactive chat, test the fast paths.

Real cost: sticker prices cluster at $3/$15, but cache-hit rate decides your actual bill. Agentic coding workloads reportedly see 90%+ cache hits, dragging effective input cost toward $0.30/M. A cheap provider without solid prefix caching can cost more than a $3/$15 provider with great caching — benchmark with your traffic before migrating.

Which provider should you pick?

Your situation	Best route
Just trying K3 today	$1 API top-up, Vercel’s $5 monthly credits, or a paid Kimi plan (the free app tier doesn’t confirm K3 access)
Reference behavior, newest features	Moonshot official API (kimi-k3)
Production app, US data boundary	Fireworks (ZDR) or Baseten
Enterprise procurement, EU residency	Nebius; Databricks if you’re a lakehouse shop
Interactive product, latency-sensitive	Modal (460 tok/s claim) or K3 Fast via Vercel AI Gateway
Already on Vercel AI SDK	Vercel AI Gateway — multi-upstream failover included
Claude Code / Cline agents	SiliconFlow (Anthropic-compatible) or official API
One key for K3 + 200 other models	OpenRouter
Side project on a tight budget	CrofAI or the budget tier — verify quality first
Compliance demands ownership	Self-host the open weights — bring ~$300K

FAQ

How many providers offer Kimi K3?

As of early August 2026, independent trackers list 10–14 API providers: Moonshot’s official API plus Together AI, Fireworks, Baseten, Modal, Nebius, SiliconFlow, Novita, DigitalOcean, Databricks, OpenRouter, Vercel AI Gateway, Cloudflare Workers AI, CrofAI, Chutes, io.net, Venice and others. The count is growing weekly as more inference clouds finish their K3 deployments; by late September 2026, OpenRouter alone listed about 20 K3 endpoints from 17 providers, and Amazon Bedrock had added K3.

What is the cheapest Kimi K3 API?

It depends on whether you optimize for input or output. As of September 25, 2026, the cheapest input rate on OpenRouter is Sail Research at about $1.35/M input and $11.50/M output, with InferenceNet close behind at about $1.40/$10.75. CrofAI still has the lowest tracked output rate at $2.00/M input and $8.00/M output. The major providers — Moonshot, Together, Fireworks, Modal, SiliconFlow — all hold the official $3 input / $0.30 cached input / $15 output line. Real cost depends more on cache-hit rates and output length than on sticker price, so a well-cached $3/$15 provider usually beats a cheap endpoint that doesn’t cache.

Which providers host Kimi K3 in the US or Europe?

Fireworks AI explicitly advertises a US-hosted K3 serverless endpoint with zero data retention, and Baseten and Together serve from US infrastructure. Vercel AI Gateway routes to US-based providers (including Baseten and Fireworks) and supports ZDR. For EU data residency, Nebius (Amsterdam) is the natural pick. Regulated teams should still verify region and retention terms in contract.

Is Kimi K3 on AWS Bedrock, Azure or Google Cloud?

Partly, as of September 2026. Kimi K3 became generally available on Amazon Bedrock on September 18, 2026, at the same $3/$15 list price on global routing. Microsoft Foundry offers it through Fireworks. Google Vertex does not carry it natively. Other routes are the official Moonshot API, third-party inference providers, gateways like OpenRouter and Vercel AI Gateway, or self-hosting the open weights.

What is Kimi K3 Fast?

Kimi K3 Fast is a lower-latency serving tier of the same model, exposed on Vercel AI Gateway as moonshotai/kimi-k3-fast (or via a speed option on the base model). It costs roughly 50% more per token in exchange for faster responses, and falls back to standard speed when the fast tier is unavailable. The underlying model, context window and quality are identical.

Which Kimi K3 provider is best for coding agents like Claude Code or Cline?

For Kimi’s own harness, Kimi Code CLI talks to the official API directly. For Claude Code, Cline and other Anthropic-format tools, SiliconFlow offers an Anthropic-compatible endpoint, and the official Moonshot API plus Fireworks both support Anthropic-style requests. Pick whichever gives you reliable cache hits — agentic coding workloads see 90%+ cache-hit rates, which is where the real savings live.

Can I try Kimi K3 for free?

Partly. The Kimi app and kimi.com still have a free tier for trying Kimi, but Moonshot’s docs list K3 access only on paid plans (Plus and up, or legacy Moderato and up) and don’t say which model free users get. Vercel AI Gateway gives unpaid teams $5 in credits every 30 days, and OpenRouter occasionally provisions zero-cost :free community routes after big launches. For guaranteed K3 access, the cheapest paths are a $1 top-up on the official API or a paid Kimi plan.

Is OpenRouter a good way to access Kimi K3?

Yes. At launch OpenRouter’s K3 listing ran on a single upstream provider, but by late September 2026 it routes across roughly 20 K3 endpoints from 17 providers, so you now get real failover as well as consolidated billing and one API key. If multi-provider routing matters, Vercel AI Gateway already routes across multiple K3 upstreams with automatic fallback. If you just want one key for everything, OpenRouter is fine.

The bottom line

Moonshot isn’t just selling a model — it’s building a distribution empire. Fourteen providers in two weeks, spanning official APIs, US inference clouds, European hosts, edge platforms, gateways and budget marketplaces, all anchored by open weights nobody can lock down. The practical takeaway for you: the access question is solved. Start on the Kimi app or official API, move to Fireworks/Baseten/Vercel when compliance or scale demands it, and leave self-hosting to the hardware reality club.

Keep reading: the Kimi K3 open weights breakdown, Kimi K3 VRAM requirements, the complete Kimi K3 guide, full specifications, and pricing head-to-heads — K3 vs Claude Fable 5, K3 vs GPT-5.6 and K3 vs GLM-5.2.