How to Use Kimi K3 Today (API + Kimi Code) — the 2.8T Open Model, Weights Dropping July 27

Read time: ~6 minutes. Key facts:

  • Kimi K3 launched July 16, 2026 — Moonshot AI’s flagship open Mixture-of-Experts model: 2.8 trillion total parameters, ~16 of 896 experts active per token (Stable LatentMoE), Kimi Delta Attention (KDA), 1M-token context, multimodal (text + image).
  • The open weights release July 27, 2026 — so you can’t run it locally yet. Until then, use the hosted paths.
  • Access today: the Moonshot API (model id kimi-k3, base URL https://api.moonshot.ai/v1, OpenAI-compatible), Kimi Code, kimi.com / the Kimi iOS app (free to try), and OpenRouter.
  • API pricing: $3.00 / 1M input, $15.00 / 1M output, with a $0.30 / 1M cache-hit input rate.

Sourcing note: architecture, capabilities, benchmarks, and the July 27 weights date are from Moonshot’s official Kimi K3 announcement; the API model id, endpoint, and pricing are cross-checked against the Moonshot API and aggregators (OpenRouter, llm-stats). The license wasn’t stated in the announcement — confirm it at the weights release. A “run Kimi K3 locally” guide follows once the weights are out. Links at the bottom.

Moonshot just dropped Kimi K3, and it’s the biggest open model yet by raw size — 2.8 trillion parameters. The catch for the “can I run it myself?” crowd: the weights don’t land until July 27, so today isn’t a local-inference story. But K3 is fully usable right now through the API and Kimi Code — and for almost everyone, that’s how you’ll want to try a 2.8T model anyway. Here’s how to access it today, and what to plan for when the weights drop.

1. What Kimi K3 is

From Moonshot’s announcement:

  • 2.8 trillion total parameters, a sparse Mixture-of-Experts with roughly 16 of 896 experts active per token (“Stable LatentMoE”) — so despite the enormous total, only a small fraction fires per token.
  • Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), Gated MLA, and a Sigmoid-Tanh Unit (SiTU) — a notably novel attention/architecture stack, not a stock transformer.
  • 1M-token context (1,048,576).
  • Multimodal — accepts text and images (the announcement also cites video understanding), aimed at long-horizon coding, knowledge work, agentic workflows, and reasoning.
  • Open weights — releasing July 27, 2026.

The “16 of 896 experts” design is the important part for later local plans: like other giant MoEs, K3 activates only a sliver of its parameters per token, so compute is modest — but you’d still need to hold a huge model in memory to serve it. (More on that when the weights land.)

2. What it can do: benchmarks and positioning

Moonshot’s reported results, framed honestly:

  • DeepSWE: 67.3 (mini-SWE-agent harness) — strong agentic-coding performance.
  • Moonshot says K3 outperforms Claude Opus 4.8 and GPT-5.5, and is competitive with Claude Fable 5 on kernel optimization.
  • It trails Claude Fable 5 and GPT-5.6 Sol overall — Moonshot is upfront that K3 isn’t the outright frontier leader.

The honest read: K3 is a genuine frontier-adjacent open model — beating last-generation flagships (Opus 4.8, GPT-5.5) and trading blows on specific tasks, while not topping the very newest closed models. For an open 2.8T model you’ll be able to self-host and fine-tune, “beats Opus 4.8, competitive on kernels, trails only the newest closed models” is a strong place to be.

3. Access today, Method 1 — kimi.com and the app (free to try)

The zero-setup path: Kimi K3 is free to try inside the Kimi iOS app and on kimi.com under standard rate limits. If you just want to feel out its coding and reasoning before wiring up an API, start here — no key, no billing.

4. Access today, Method 2 — the Moonshot API (OpenAI-compatible)

For building, use the Moonshot API. It’s OpenAI-compatible, so you point an existing OpenAI SDK at Moonshot’s base URL and swap the model name:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_MOONSHOT_API_KEY",
    base_url="https://api.moonshot.ai/v1",   # Moonshot endpoint
)

resp = client.chat.completions.create(
    model="kimi-k3",
    messages=[
        {"role": "user", "content": "Refactor this function and explain the change."},
    ],
)
print(resp.choices[0].message.content)

The essentials:

  • Model id: kimi-k3. Base URL: https://api.moonshot.ai/v1.
  • Pricing: $3.00 per 1M input tokens, $15.00 per 1M output tokens, and a $0.30 per 1M cache-hit input rate — so structuring requests to reuse a cached prefix (a stable system prompt / context) cuts input cost roughly 10×. With a 1M context and per-token output cost, watch your output length on long agentic loops.
  • Because it’s OpenAI-compatible, most existing tooling works by changing just the base URL and model name.

It’s also available on OpenRouter (moonshotai/kimi-k3) if you’d rather route through an aggregator and avoid a separate account.

5. Access today, Method 3 — Kimi Code

Moonshot launched Kimi Code alongside the model on July 16. If you want K3 driving an agentic coding CLI (the use case its DeepSWE score targets), Kimi Code is the first-party way in — point it at your repo and let K3 run the tool-using loop, rather than wiring the agent yourself against the raw API.

6. Running it locally: what to expect July 27

This is the honest part: you cannot run Kimi K3 locally today — there are no open weights, no HuggingFace repo, and no GGUF until the July 27 release. When the weights land:

  • vLLM support is coming with the release — Moonshot contributed a KDA prefix-caching implementation “to be released alongside the model,” so vLLM will be a day-one serving path.
  • Expect the same honest-hardware reality as other giant MoEs: a 2.8T model, even quantized, needs serious memory to hold — this will be datacenter / very-high-RAM territory, not a laptop. (For the shape of that math, see how it played out for Kimi K2.7 Code and Inkling.)
  • License: not stated in the announcement — confirm the exact terms when the weights drop before building on them commercially.

We’ll publish a “run Kimi K3 locally” guide once the weights and the GGUF/vLLM recipes are actually out and verifiable — no point writing local commands against a model you can’t yet download.

The takeaway

Kimi K3 is Moonshot’s new 2.8-trillion-parameter open MoE — Kimi Delta Attention, 1M context, multimodal, and strong enough to beat Opus 4.8 and GPT-5.5 on Moonshot’s benchmarks (DeepSWE 67.3). But the open weights don’t drop until July 27, so today it’s an access story, not a local one: try it free on kimi.com / the iOS app, build against the API (kimi-k3 at https://api.moonshot.ai/v1, OpenAI-compatible, $3/$15 per 1M, $0.30 cache-hit), or run agentic coding through Kimi Code. Use the cache-hit pricing by keeping a stable prompt prefix. When the weights land on the 27th — vLLM with KDA support in tow — the local guide follows.

For the local-run playbook on models whose weights are already out, see Kimi K2.7 Code locally, Inkling locally, and GLM-5.2 locally.

Sources

  • Kimi K3: Open Frontier Intelligence — Moonshot AI — 2.8T total params, ~16 of 896 experts active (Stable LatentMoE), Kimi Delta Attention (KDA) / AttnRes / Gated MLA / SiTU, 1M context, multimodal (text/images/video), benchmarks (DeepSWE 67.3; beats Opus 4.8 and GPT-5.5; competitive with Fable 5 on kernel optimization; trails Fable 5 and GPT-5.6 Sol overall), weights releasing July 27 2026, vLLM KDA prefix-caching to ship with the model
  • Moonshot API — model id kimi-k3, base URL https://api.moonshot.ai/v1, OpenAI-compatible; pricing $3.00 / 1M input, $15.00 / 1M output, $0.30 / 1M cache-hit input; free to try on kimi.com and the Kimi iOS app; Kimi Code launched July 16, 2026; also on OpenRouter as moonshotai/kimi-k3
  • License not disclosed in the announcement — confirm at the July 27 weights release. Benchmarks are Moonshot’s own reported results. Verified July 17, 2026 — a local-run guide will follow once the weights are published.