Where to Access Kimi K3: Every Provider and Price Compared (2026)

Last updated: October 6, 2026. Added the GitHub Copilot route (from the August 6 changelog) and related internal links. The OpenRouter route was last re-checked on September 28; other provider rows are the July 27 snapshot.

You can access Kimi K3 through Moonshot's first-party API, inference clouds such as Together, Fireworks, Baseten, and Modal, gateways such as OpenRouter and Vercel AI Gateway, or by self-hosting the open weights from Hugging Face. Moonshot, Together, Fireworks, Modal, and several other hosts list the same rate of $3 input, $0.30 cached input, and $15 output per million tokens. OpenRouter now lists 19 K3 endpoints, some cheaper and some pricier than that, so the right route depends on billing model, data boundary, and agent harness rather than whether K3 is available.

Referral disclosure: The OpenCode Go and RunPod links in this guide are Developers Digest referral links. You may receive a signup benefit, and Developers Digest may receive account credits or commission. Other outbound links use utm_source=developersdigest for attribution only and are not affiliate links.

Quick navigation#

• Provider details: Moonshot, open weights, Together AI, Fireworks AI, Baseten, Modal, Vercel AI Gateway, Cloudflare Workers AI, RunPod, SiliconFlow, and OpenRouter and OpenCode Go.
• Research context: provider launch posts, Devin availability, recommendations, and sources.

Kimi K3 access at a glance#

Prices are a July 27 snapshot, except OpenRouter, which was re-checked on September 28. Provider rates, cache treatment, regions, and capacity can change without the model ID changing.

Provider launch posts and implementation PRs#

Several providers published useful launch-day technical material rather than only adding a model card:

Together's comparison numbers are Together's own evaluation, and Modal's speed figures come from Modal. They are useful implementation evidence, not neutral cross-provider benchmarks.

Moonshot API, Kimi Code, and Kimi Agent#

Moonshot is the reference route. The API uses model="kimi-k3" with the OpenAI-compatible base URL https://api.moonshot.ai/v1. K3 unlocks after a successful top-up of at least $1.

The first-party API supports the full 1,048,576-token context, text and image input, structured output, tool choice, dynamic tool loading, automatic prompt caching, and low, high, or max reasoning effort. The published rate is $3 per million fresh input tokens, $0.30 per million cached input tokens, and $15 per million output tokens.

Kimi Code is the better first-party route when you want Moonshot's terminal agent instead of a raw API. Kimi chat, Agent, Work, and Code share the membership credit system.

The Kimi K3 weights are live#

Moonshot released the full model weights on Hugging Face on July 27. The model card documents vLLM, SGLang, TokenSpeed, Transformers, and Docker Model Runner paths.

Our weights release write-up covers the community reaction. Open weights do not make K3 laptop friendly. K3 has 2.8 trillion total parameters, 104 billion active parameters, and native MXFP4 weights. Baseten says the MXFP4 files exceed 1.4 TB and that one production replica uses eight NVIDIA GB300 GPUs.

The Kimi K3 License permits use, modification, deployment, fine-tuning, derivatives, distribution, and sale, but it is not plain MIT. Model-as-a-service businesses over the license's revenue threshold need a separate Moonshot agreement, and large commercial products can trigger an attribution requirement. Read the Kimi K3 License before building a hosted service.

Together AI#

Together exposes moonshotai/Kimi-K3 through serverless, dedicated, and provisioned-throughput options. Its model page lists native vision, the 1M-token context, and the same $3 input, $0.30 cached input, and $15 output rates as Moonshot.

Choose Together when you want a managed API now with a straightforward path to reserved capacity later.

Fireworks AI#

Fireworks exposes K3 as accounts/fireworks/models/kimi-k3. It supports serverless inference, on-demand dedicated GPUs, image input, function calling, and fine-tuning.

The differentiator is operational control. Fireworks says its K3 serverless endpoint is US hosted with zero data retention and offers a path from shared inference into dedicated capacity and tuning.

Baseten#

Baseten's K3 Model API supports image input and the full 1M-token context through an OpenAI-compatible endpoint. Its day-0 article is the clearest infrastructure explanation in this provider set, including validation with the Kimi Vendor Verifier, vLLM, and SGLang.

Baseten does not publish a simple token rate on the public model page. Check the console or request a quote before comparing it with per-token providers.

Modal#

Modal offers an OpenAI-compatible Shared API and a dedicated Auto Endpoint. Its model page lists $3 per million prompt tokens, $0.30 per million cached prompt tokens, and $15 per million completion or reasoning tokens.

Modal says its shared endpoint reaches 460 output tokens per second using its DFlash speculator. Treat that as a provider measurement, but consider Modal when interactive speed and a path to programmable GPU infrastructure matter.

Vercel AI Gateway#

Vercel AI Gateway exposes K3 as moonshotai/kimi-k3 through Moonshot AI and Novita AI. The model page shows live latency, throughput, and uptime data and lists the same $3 input, $0.30 cache-read, and $15 output rates. Unpaid teams receive $5 in AI Gateway credits every 30 days.

This is the cleanest route for applications already using the Vercel AI SDK. The merged AI SDK implementation adds K3 to both the Moonshot provider and AI Gateway, including the reasoningEffort option.

Cloudflare Workers AI#

Cloudflare lists K3 as moonshotai/kimi-k3 in Workers AI with the full context window and an OpenAI-compatible chat-completions format. Public documentation does not expose a fixed K3 token price, so check the Cloudflare dashboard.

Choose Cloudflare when inference belongs inside an existing Workers, observability, and billing stack.

RunPod#

RunPod exposes K3 through its shared moonshot-kimi public endpoint. Set model="kimi-k3" in the request body. The same endpoint also serves K2.6 and K2.7 Code.

RunPod's public page lists $15 per million tokens without separating input, cached input, and output. The provider is also the strongest public referral option in this group.

Create a RunPod account through the Developers Digest referral link. New referred users can receive a one-time credit after their first qualifying deposit. Developers Digest earns credits on qualifying Pod and Serverless spend and can unlock the cash affiliate tier after 25 paying referrals.

SiliconFlow#

SiliconFlow lists K3 at $3 per million input tokens, $0.30 per million cache reads, and $15 per million output tokens. It supports OpenAI-compatible and Anthropic-compatible request formats.

Its launch article documents K3 with Claude Code, Cline, Hermes Agent, and OpenCode. That makes SiliconFlow a useful compatibility layer, not proof that those clients use K3 as their default model.

OpenRouter and OpenCode Go#

OpenRouter exposes moonshotai/kimi-k3 behind one key and billing layer. On September 28 its endpoint list showed 19 endpoints, including Moonshot AI, Together, Fireworks, Baseten, Modal, DeepInfra, and others, so it now gives real provider failover. Input prices ranged from $0.88 to $3.45 per million tokens and output from $7.86 to $17.25. The cheapest endpoints are fp4 or fp8 quantized (some list quantization as unknown), while Moonshot's own endpoint is native mxfp4 at $3 and $15, so test quality before optimizing on price.

OpenCode Go is an agent subscription rather than a raw per-token API. OpenCode Go through the Developers Digest referral link is $5 for the first month under the current offer, then $10 per month. It is the lowest-friction way to evaluate K3 inside a coding-agent loop.

Can you use Kimi K3 in Devin?#

No verified K3 access route exists in Devin. Cognition's enterprise deployment documentation describes Devin as a compound AI system and says it does not support third-party LLM API keys. Devin's public selector does not expose K3 as a user-selectable model.

That does not prove Cognition never uses Moonshot technology internally. It means a developer cannot currently select K3 in Devin or bring a K3 provider key.

Which Kimi K3 route should you pick?#

• Fastest agent-first trial: OpenCode Go.
• Direct API and reference behavior: Moonshot.
• Managed API with dedicated capacity: Together AI.
• US hosting, zero retention, fine-tuning, or dedicated GPUs: Fireworks.
• Managed serving with infrastructure detail: Baseten.
• High interactive speed and programmable infrastructure: Modal.
• Existing Vercel AI SDK application: Vercel AI Gateway.
• Existing Cloudflare application: Workers AI.
• Shared endpoint plus a real public referral program: RunPod.
• OpenAI and Anthropic client compatibility: SiliconFlow.
• One key across many model vendors: OpenRouter.
• Already paying for Copilot: pick K3 in the model picker; Business and Enterprise admins must enable the policy first, per the rollout write-up.
• Self-hosting or research: Hugging Face weights, after reviewing the license and infrastructure requirements.

Referral and partner options#

Two links are usable now:

• OpenCode Go provides the current Developers Digest first-month offer.
• RunPod has a public referral program and a path to cash affiliate status after 25 paying referrals.

Kimi also has a credit-only invitation campaign. Fireworks, Modal, and Vercel accept partner applications, but none provides a public creator commission schedule on its application page. Baseten's referral or resale application is currently lower priority because it publishes no concrete creator offer.

FAQ#

Are the Kimi K3 weights available?#

Yes. Moonshot released the model card, weights, serving guidance, and Kimi K3 License on Hugging Face on July 27, 2026.

What is the cheapest Kimi K3 API?#

Moonshot, Together, Fireworks, Modal, and SiliconFlow advertise the same $3 input and $15 output rate per million tokens, with $0.30 cached input where published. On OpenRouter, some endpoints list input as low as $0.88 and output as low as $7.86, as of September 28. Real cost depends on quantization, cache treatment, output length, rate limits, and provider fees.

Which provider hosts Kimi K3 in the United States?#

Fireworks explicitly advertises a US-hosted K3 serverless endpoint with zero data retention. Verify region and contract terms for regulated workloads.

Can I run Kimi K3 locally?#

The weights are downloadable, but K3 is a cluster-scale model. The MXFP4 files exceed 1.4 TB, and a production replica can require eight GB300 GPUs. It is not a practical laptop model.

Can I select Kimi K3 in Devin?#

No verified option exists. Devin does not accept third-party model API keys, and Cognition has not published a user-selectable K3 integration.

Can I earn referral revenue from a K3 provider?#

RunPod has the clearest public program, including a path to 10% cash commission after 25 paying referrals. OpenCode Go already has a Developers Digest referral link. Other providers require partner outreach or offer non-cash credits.

Continue Reading#

• Open-Weight Model Hosting Prices - every provider's Kimi K3 per-token price, context and variant side by side, each value quoted from the provider's page
• Where to Access AI Models in 2026 - the hub covering access routes, free tiers, and prices for every major model, not just K3
• Kimi K3 Developer Guide: What the 2.8T Open Model Changes
• Kimi K3 vs K2.7: Is the Upgrade Worth It for Coding?
• Kimi K3 Drops: Moonshot's 2.8T Frontier Model
• OpenCode Developer Guide 2026
• Model Routers and the Optionality Advantage
• Kimi Code vs Claude Code - side-by-side for the two agent-first clients
• Kimi K3 Is GA in GitHub Copilot - rollout, pricing and the admin policy for the Copilot route
• Kimi K3 Weights Land on HuggingFace - the open-weights release and license reaction

Sources#

• Moonshot Kimi K3 technical blog - fetched July 27, 2026
• Kimi K3 API quickstart - fetched July 27, 2026
• Kimi K3 Hugging Face model card - fetched July 27, 2026
• Together AI Kimi K3 model page - fetched July 27, 2026
• Fireworks Kimi K3 model page - fetched July 27, 2026
• Fireworks Kimi K3 launch article - fetched July 27, 2026
• Baseten Kimi K3 model page - fetched July 27, 2026
• Baseten day-0 Kimi K3 article - fetched July 27, 2026
• Modal Kimi K3 launch article - fetched July 27, 2026
• Vercel AI Gateway Kimi K3 model page - fetched July 27, 2026
• Cloudflare Workers AI Kimi K3 docs - fetched July 27, 2026
• RunPod Moonshot Kimi endpoint docs - fetched July 27, 2026
• SiliconFlow Kimi K3 launch article - fetched July 27, 2026
• OpenRouter Kimi K3 model page and its endpoints API - fetched September 28, 2026