By the end of this article, you'll know:
- what Kimi K3 actually is, the architecture, the scale, and what changed from the K2 family
- how to run it today, free in the browser, through the API, or wired into Claude Code, Cursor, Cline, and RooCode
- which exact model ID and settings unlock the full 1 million token context
- how K3 stacks up on price and benchmarks against DeepSeek V4, Qwen3.7 Max, GLM-5.2, and its own sibling K2.7 Code
- whether switching today makes sense, or whether you should wait
Kimi K3 just dropped, and Moonshot means business
Moonshot AI released Kimi K3 on July 16, 2026, timed just ahead of the World Artificial Intelligence Conference in Shanghai. The Beijing lab, backed by Alibaba, has spent the past 18 months watching DeepSeek erode its market position. K3 is the comeback attempt, and the early numbers back it up.
The headline result: K3 debuted at number one on Arena.ai's Frontend Code Arena with a score of 1,679, ahead of Claude Fable 5 at 1,631 and GPT-5.6 Sol at 1,618. Its predecessor, K2.6, sat 18th on that same board. That's a 17 spot jump in one release, and it's independently measured, not a number Moonshot invented itself.
This isn't Moonshot's first time in Western production stacks either. Cursor built its Composer 2 model starting from a Kimi K2.5 base. DoorDash's CTO has said the company routes lower level work to Kimi K2.6. Thinking Machines used K2.5 to help generate early post-training data for Inkling, its own model released just a day before K3. So when a new Kimi flagship lands, it lands somewhere developers already have skin in the game.
For developers, the practical story matters more than the leaderboard. K3 is open weight, speaks the OpenAI SDK, and drops into tools you already use. It's also, for the first time in Kimi's history, priced like a frontier model instead of a budget one. That changes the calculus on when it's actually worth reaching for.
What Kimi K3 actually is
K3 is a sparse Mixture-of-Experts model with 2.8 trillion total parameters, Moonshot's largest release yet and, by the company's own claim, the largest open-weight model anyone has shipped. Early leaks in the days before launch put the figure closer to 2.5 trillion. The confirmed number, repeated in Moonshot's own materials and independent trackers like Artificial Analysis and LLM Stats, is 2.8 trillion. Community estimates based on the "2.8T-A50B" shorthand circulating online put active parameters at roughly 50 billion per token, though Moonshot has not officially published that number, so treat it as directionally right rather than confirmed.
The architecture runs on what Moonshot calls Stable LatentMoE, routing each token to 16 of 896 experts. Two new mechanisms sit on top of that base. Kimi Delta Attention is a hybrid linear attention design that Moonshot says enables up to 6.3 times faster decoding at million token context. Attention Residuals selectively retrieve representations across depth instead of accumulating them uniformly through every layer, which Moonshot claims delivers roughly 25 percent higher training efficiency for under 2 percent extra cost. Both figures are vendor reported and pending independent reproduction.
K3 ships with a 1,048,576 token context window, native image and video understanding, and thinking mode on by default. That's a real jump from the K2 line. K2 launched in July 2025 at 1 trillion total parameters and 32 billion active. K2.5 followed in January 2026 with multimodal support. K2.6 arrived in April. K2.7 Code shipped June 12 as a coding specialist, built on that same 1T/32B base with a 256K window. K3 isn't a point release. It's a new architecture at nearly three times the scale of anything Moonshot has shipped before, and the first Kimi model to top an arena leaderboard outright rather than just leading among open models.
How to use Kimi K3
Try it free. The fastest path is the Kimi app, on the web at kimi.com or on iOS. K3 Max is available to any free account under standard rate limits, no API key needed. The full 1 million token window is gated to paid membership tiers though, it isn't handed out by default on the free plan.
Set up the API. Sign up at platform.kimi.ai, go to API Keys, and generate one. The API speaks the OpenAI SDK directly, so integration is a base URL swap:
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["MOONSHOT_API_KEY"],
base_url="https://api.moonshot.ai/v1",
)
completion = client.chat.completions.create(
model="kimi-k3",
messages=[{"role": "user", "content": "Introduce Kimi K3 in one sentence."}],
)
print(completion.choices[0].message.content)
The model ID on this general API is kimi-k3. Default max output is 131,072 tokens, configurable up to the full window. Thinking mode is always on, and right now only "max" reasoning effort is supported, low and high effort are coming later.
Use the CLI or IDE extension. Kimi Code is Moonshot's terminal and VS Code tool built specifically for coding agent workflows, and it's a separate product with its own endpoint. It offers two models: K3 as the flagship, model ID k3, and K2.7 Code as the mature coding specialist, model ID kimi-for-coding or kimi-for-coding-highspeed for a faster tier at the same ability level. The Kimi Code base URL is https://api.kimi.com/coding/v1. Many third-party tools default to a smaller context window than the model supports, so set it manually to 1048576 to actually get the full 1M.
Plug it into tools you already use. Moonshot's own docs list Claude Code, Cline, RooCode, and Codex as supported coding agent harnesses. For Claude Code specifically, point its environment at Kimi's coding endpoint:
export ANTHROPIC_BASE_URL=https://api.kimi.com/coding/
export ANTHROPIC_API_KEY=Your API Key
export ANTHROPIC_MODEL="k3[1m]"
export ANTHROPIC_DEFAULT_SONNET_MODEL=$ANTHROPIC_MODEL
export ANTHROPIC_DEFAULT_OPUS_MODEL=$ANTHROPIC_MODEL
export CLAUDE_CODE_MAX_CONTEXT_TOKENS=1048576
export CLAUDE_CODE_EFFORT_LEVEL=max
claude
Cline and RooCode work the same way in principle, point the custom OpenAI-compatible provider setting at Moonshot's base URL and drop in the model ID. Cursor has no native K3 listing yet at launch, there's an open request for it on Cursor's own community forum, though given Composer 2's own K2.5 lineage, that gap likely won't last long.
Pricing
K3 costs $3.00 per million input tokens on a cache miss, $0.30 per million on a cache hit, and $15.00 per million output tokens, flat across the entire 1M context with no long-prompt surcharge. That's the rate on Moonshot's own pricing page, and it matches independent trackers including Artificial Analysis and OpenRouter. One Axios piece cited a blended figure near $12 per million tokens, but that reads as an outlier against the far more consistently reported $3/$15 sticker price everywhere else, including Moonshot's own docs. Cache hit rate is the biggest lever on real cost: Moonshot reports over 90 percent cache hit rates on coding workloads, and agent loops that resend repo context benefit heavily from that 10x discount. Output verbosity is the other lever. K3 uses roughly 21 percent fewer output tokens than K2.6 did on Artificial Analysis's Intelligence Index run, which works in your favor even at the higher per-token sticker price.
how kimi k3 compares to other top AI models:
• kimi k3 - #1 on the frontend arena leaderboard (elo 1,679). 1M context, $3 input / $15 output, open weights expected July 27.
• claude fable 5 - #2 (elo 1,631). output costs $50/M, closed-source.
• gpt-5.6 sol - #3 (elo 1,618). 1.05M context, $5 input / $30 output, closed-source.
• glm-5.2 - #4 (elo 1,587). 1M context, $1.40 input / $4.40 output, open-source (MIT).
• deepseek v4 pro - 1M context, $0.44 input / $0.87 output, open-source (MIT) and currently the cheapest option.
• qwen3.7 max - 1M context, $2.50 input / $7.50 output, closed-source.
• kimi k2.7 code - 256K context, $0.95 input / $4 output, open-source (Modified MIT).
kimi k3 is currently the strongest model on the leaderboard, deepseek is the best value for money, and claude/gpt remain the premium closed-source options.
Against its direct open-weight rivals, K3 is the expensive option.
DeepSeek V4 Pro undercuts it roughly seven times over on input and seventeen times over on output, and GLM-5.2 comes in around a third to half the price. Neither matches K3's Frontend Arena score, but both sit close enough on general capability that price becomes the deciding factor for high volume workloads.
Qwen's own flagship, notably, went proprietary with the 3.7 Max release, a break from Alibaba's earlier open-weight pattern, and its Artificial Analysis Intelligence Index score of 46 trails K3's 57 by a wide margin.
Against closed frontier models, K3 leads on the one benchmark Moonshot chose to headline: frontend code generation by human preference, beating Claude Fable 5 and GPT-5.6 Sol outright and topping six of seven sub-domains. On broader intelligence it still trails both.
Artificial Analysis puts K3 at 57 against roughly 59 to 60 for GPT-5.6 Sol and Claude Fable 5, landing it in the same tier as Claude Opus 4.8 and GPT-5.5 instead.
Cost per task tells a friendlier story: K3 runs about $0.94 per task on the Intelligence Index, similar to GPT-5.6 Sol's $1.04, and about half of Opus 4.8's $1.80.
Reality check
Confirmed and independently verified: 2.8T parameters, the 1M context window, the KDA and Attention Residuals architecture, the $3/$15 pricing, and the Frontend Code Arena result, which came from Arena.ai's own voting data, not from Moonshot.
Still rolling out: full open weights, promised by July 27 and not yet on Hugging Face at time of writing. Broader reasoning effort levels beyond max. Wider third-party IDE support.
Where lab claims and independent numbers diverge: Moonshot's own benchmark table mixes results run through different agent harnesses, Kimi Code, Claude Code, Codex, and mini-SWE-agent, which makes side by side comparison harder than the headline chart suggests. On ProgramBench specifically, the benchmark's own author flagged that Kimi used an averaging implementation percentage metric rather than counting fully working programs, a choice that can overstate real world usefulness. The 6.3x decoding speedup and 25 percent training efficiency figures are both vendor reported and unverified independently as of this writing.
The clearest real limitation is hallucination. On Artificial Analysis's AA-Omniscience Index, K3's raw accuracy improved from 33 to 46 percent over K2.6, but its hallucination rate climbed from 39 to 51 percent in the same run. It gets more questions right and fabricates more answers at the same time, a real tradeoff worth knowing before pointing it at anything factual and unsupervised. Output speed also runs slower than the median model in its price tier, and self-hosting a 2.8 trillion parameter model, once weights are actually out, is realistically an option for well funded teams only.
My final verdict
Use Kimi K3 today if you need frontier-level frontend and UI generation, you're already working inside Claude Code, Cline, or RooCode and want a capable drop-in option, or you specifically need the combination of a 1M context window and native vision without paying full Western-flagship prices.
Wait if budget is your main constraint. DeepSeek V4 Pro and GLM-5.2 deliver most of the capability for a fraction of the cost, and both already have open weights in hand. Wait if you need to self-host right now, since K3's weights aren't public yet. Wait if hallucination rate matters more to your use case than raw accuracy. And if K2.7 Code already handles your coding workload well, there's no urgent reason to move. K3 is the bigger, pricier, more general sibling, not a strict upgrade for every task.
