Back to blog

Kimi K3 and China's Open-Weight Model Wave

Kimi K3 is the largest open-weight model ever released — 2.8T parameters, a 1M-token context, and benchmark scores next to Claude Opus 4.8. Here is what it costs, how it compares to GLM-5.2, DeepSeek V4 Pro, and MiniMax M3, and how to run all of them through one API with LLM Gateway — on a flat-rate DevPass plan or pay-as-you-go credits.

Glossy circuit board with a glowing central chip radiating light traces, representing Kimi K3 and open-weight models routing through one gateway

Moonshot AI released Kimi K3 on July 16, and the benchmarks put an open-weight model next to the best closed ones for the first time. The catch is access. K3 sits on Moonshot's platform, GLM-5.2 on Z.ai's, DeepSeek V4 Pro on DeepSeek's, MiniMax M3 on MiniMax's — four accounts, four billing relationships, four API dashboards, all before you have written a line of code.

LLM Gateway routes every one of them through a single OpenAI-compatible endpoint. One key, one bill, and a switch between Kimi K3 and any of 200+ models is a one-word change to your request.

What is Kimi K3?

Kimi K3 is Moonshot AI's flagship model for long-horizon coding and agentic work. At 2.8 trillion parameters — a mixture-of-experts design that activates 16 of its 896 experts per token — it is the largest open-weight model announced to date. Moonshot has committed to publishing the full weights by July 27, 2026.

The specs that matter in practice:

  • 1M-token context window (1,048,576 tokens), with output configurable up to the same 1M — enough to hold a large repository plus its docs in a single request
  • Always-on reasoning — K3 thinks before every answer; there is no non-thinking mode
  • Vision, tool calls, and JSON output supported out of the box
  • Prompt caching at a 90% discount on repeated input

Early results back up the size. K3 ranks fourth of 189 models on the Artificial Analysis Intelligence Index — level with Claude Opus 4.8 and GPT-5.5 — and took first place in Arena's Frontend Code evaluation in blind developer testing. It posted 93.5% on GPQA Diamond and 88.3% on Terminal-Bench 2.1, the strongest open-weight results published on both at release.

Kimi K3 pricing

Through LLM Gateway you pay Moonshot's published per-token rates:

Tokens Price per million
Input $3.00
Cached input $0.30
Output $15.00

The cached-input rate is the number to watch. Coding agents re-send the same system prompt, file context, and conversation history on every step, so in a long agent session most of your input tokens are cache hits billed at a tenth of the fresh rate.

The open-weight wave K3 leads

K3 is the headline, but 2026 has produced a full roster of frontier-class open-weight models from Chinese labs — all MIT-licensed or close to it, all available through LLM Gateway at provider list prices:

Model Lab Released Context Input $/M Output $/M
Kimi K3 Moonshot AI Jul 2026 1M $3.00 $15.00
GLM-5.2 Z.ai Jun 2026 1M $1.40 $4.40
MiniMax M3 MiniMax Jun 2026 512K $0.60 $2.40
DeepSeek V4 Pro DeepSeek Apr 2026 1M $0.435 $0.87

Each fills a different slot. GLM-5.2 (744B, MIT license) was the top-ranked open model before K3 arrived — its output costs under a third of K3's rate ($4.40 vs $15.00 per million) and its input about 47% ($1.40 vs $3.00). DeepSeek V4 Pro (1.6T total, 49B active, MIT license) delivers a 1M-token context at under a dollar per million output tokens. MiniMax M3 is the multimodal option, taking image input alongside text. And if you want a zero-cost entry point, glm-4.7-flash-free runs through the same endpoint for $0.

One endpoint for all of them

LLM Gateway speaks the OpenAI API, so anything that can call OpenAI can call Kimi K3:

1curl https://api.llmgateway.io/v1/chat/completions \2  -H "Authorization: Bearer $LLM_GATEWAY_API_KEY" \3  -H "Content-Type: application/json" \4  -d '{5    "model": "kimi-k3",6    "messages": [{"role": "user", "content": "Refactor this module to remove the circular import."}]7  }'

Swapping to GLM-5.2 or DeepSeek V4 Pro is a change to the model field — no new SDK, no new key. If you need a specific upstream, pin it with the provider prefix (moonshot/kimi-k3); otherwise the gateway routes to a healthy provider and fails over automatically if one goes down. Kimi K3 is served by both Moonshot and Novita, so a single-provider outage does not take your app with it.

Every request lands in the same dashboard: per-model cost, token counts, cache-hit rates, and latency, across all providers.

Run Kimi K3 on a flat rate with DevPass

If K3 is headed for your coding agent, the cheapest way to run it hard is DevPass — a flat monthly plan that works with Claude Code, OpenCode, Cline, Cursor (chat/plan mode only), and anything else OpenAI- or Anthropic-compatible:

Plan Price Model usage included
Lite $29/mo ~$87 at provider rates
Pro $79/mo ~$237
Max $179/mo ~$537

Every dollar of allowance is metered at the provider rates above, so on Pro you get roughly $237 of model usage for $79 — and a runaway agent loop cannot produce a surprise invoice.

One gating detail to know: at $15 per million output tokens, Kimi K3 counts as a premium-tier model on DevPass, so it draws from a weekly premium allowance — roughly $10 per week of premium usage on Lite, $36 on Pro, and $97 on Max, on top of your monthly pool. GLM-5.2, MiniMax M3, and DeepSeek V4 Pro are all standard-tier: no weekly cap, usable until your monthly allowance runs out. A workable pattern is K3 for the hardest tasks and GLM-5.2 or DeepSeek V4 Pro for the bulk of the loop. If you burn through the premium allowance mid-week, a Reset Pass restores it instantly — Pro includes one per cycle and Max includes two.

Or pay as you go

For production apps, PAYG credits skip the subscription entirely. Top up from $10, pay each provider's published per-token price as you go, and LLM Gateway adds a flat 5% platform fee at top-up — the per-token rates themselves are pass-through. You keep the same single endpoint, automatic failover, prompt caching, and cost dashboard, metered to the token.

The two models work together: DevPass for your own coding agents, PAYG for the product you ship.

Frequently Asked Questions

Is Kimi K3 open source?

Kimi K3 is an open-weight model. Moonshot AI announced it on July 16, 2026 and has committed to releasing the full 2.8T-parameter weights by July 27, 2026; the weights license has not been announced yet. The API is live now — self-hosting will be possible once the weights publish, though a 2.8T mixture-of-experts model demands serious hardware, which is why most teams will run it through an API.

How much does Kimi K3 cost?

Through LLM Gateway, Kimi K3 costs $3.00 per million input tokens, $0.30 per million cached input tokens, and $15.00 per million output tokens — Moonshot's published rates. On a DevPass plan it is included in the flat monthly price as a premium-tier model.

What is Kimi K3's context window?

1,048,576 tokens — a full 1M-token context, with output defaulting to 131K tokens and configurable up to the same 1M. The same headline context as GLM-5.2 and DeepSeek V4 Pro.

Can I use Kimi K3 with Claude Code or Cursor?

Yes. LLM Gateway exposes OpenAI- and Anthropic-compatible endpoints, so point your tool's base URL at https://api.llmgateway.io/v1 with your key and select kimi-k3 as the model. In Claude Code and Cline the full agent loop routes through LLM Gateway; Cursor honors an external endpoint only in its chat/plan panel — Composer, inline edit, and autocomplete stay on Cursor's own backend. With DevPass this is covered by the flat monthly rate. Step-by-step setup for each tool: How to Use Kimi K3 with Claude Code, Cursor, and Cline.

Getting started