How we assess: AITrendyReview rates AI tools on what they cost to actually run, whether the benchmarks survive real use, and who each model genuinely suits. Here’s our editorial policy.
Last updated: September 2026. Kimi K3 is new and fast-moving; benchmark rankings and pricing shift weekly, so I’ve dated every figure below.
The short answer: what is Kimi K3 and does it matter?
Kimi K3 is Moonshot AI’s new 2.8-trillion-parameter open-weight model — the largest open model ever released — and yes, it matters, because it’s the first freely downloadable model that trades blows with the closed frontier on coding. It topped the Frontend Code Arena at launch, ahead of Claude Fable 5, and it ships with a 1-million-token context window and native vision. If you build software, Kimi K3 is the most interesting thing to happen to open weights in 2026.
But “open” here comes with a giant asterisk: at roughly 1.4TB of weights, you can’t run Kimi K3 on your own hardware without a small server farm. For almost everyone, “using Kimi K3” means paying for an API — and there, the maths gets more interesting than the hype suggests. Let me separate what’s confirmed from what’s marketing.
- 2.8T parameters, Mixture-of-Experts: 896 experts, only ~16 active per token.
- 1M-token context (1,048,576) with native vision.
- Ranked #1 in Frontend Code Arena at launch (1,679), ahead of Claude Fable 5.
- Weights released July 27, 2026 — but ~1.4TB means you can’t self-host on consumer hardware.
- API pricing (Sept 2026): $3 / 1M input, $15 / 1M output, $0.30 cache-hit input.
- The license was not fully published at reveal — read it before you build on it.
What Kimi K3 actually is
Kimi K3 is a frontier-scale large language model from Moonshot AI, the Beijing lab behind the Kimi assistant. Its defining trait is scale-with-openness: 2.8 trillion total parameters released as downloadable weights, which Moonshot bills as the world’s first open 3T-class system. That’s roughly an order of magnitude larger than the open models most people have actually run.
It’s a Mixture-of-Experts (MoE) design, and this is the key to how it stays usable. Of its 896 experts, only about 16 fire on any given token — roughly 1.8% of the network — so the compute per token is a fraction of what the headline parameter count implies. Moonshot pairs that with a 1-million-token context window (precisely 1,048,576 tokens), native vision, and two architectural tricks it credits for a claimed 2.5x jump in scaling efficiency over Kimi K2: Kimi Delta Attention, a hybrid linear-attention scheme, and Attention Residuals. The weights ship in MXFP4 with MXFP8 activations, chosen so more hardware can load them.
In plain terms: it’s a genuinely massive open model engineered so that running it is merely very hard, rather than impossible. If you follow the closed frontier, our reviews of GPT-5.4 and Claude Opus are the natural points of comparison.
Who Kimi K3 is for
Kimi K3 is aimed at three groups. Developers and AI engineers who want frontier coding ability without being locked into one US vendor. Enterprises and researchers who need to run a top-tier model on their own infrastructure for privacy or sovereignty reasons — and who have the hardware budget to do it. And builders who simply want the cheapest capable model on a per-token basis via API and don’t care where it’s hosted.
It is emphatically not for the hobbyist hoping to run a frontier model on a gaming PC. We’ll get to why.
Kimi K3 specs and pricing at a glance
| Attribute | Detail (as of Sept 2026) |
|---|---|
| Total parameters | 2.8 trillion (largest open-weight model to date) |
| Architecture | Mixture-of-Experts, 896 experts, ~16 active per token |
| Context window | 1,000,000 tokens (1,048,576), native vision |
| Quantization | MXFP4 weights, MXFP8 activations |
| Weights size | ~1.4TB (MXFP4) |
| Weights release | July 27, 2026 |
| API input price | $3.00 / 1M tokens (cache miss); $0.30 / 1M (cache hit) |
| API output price | $15.00 / 1M tokens |
| Self-host requirement | 64+ accelerators (supernode); not runnable on a single H100/H200/B200 |
| License | Not fully published at reveal — verify terms before commercial use |
The benchmarks, honestly read
Frontend coding: the real headline
The claim that made everyone look up: Kimi K3 took first place in the Frontend Code Arena (Arena WebDev) at launch with a preliminary score of 1,679, edging out Anthropic’s Claude Fable 5. Reports also had it beating Claude Opus 4.8 and GPT-5.5 across a spread of coding and agentic tasks. For an open model, being the one to beat at frontend generation — not merely “competitive” — is a first.
Verdict: the coding story is real and it’s the reason to care.
[ADD YOUR EXPERIENCE: run Kimi K3 on a real front-end task you know well and note how its output compares to Claude or GPT in your own workflow.]
Overall intelligence: strong, not #1
Away from its home turf, the picture is more measured. On the Artificial Analysis Intelligence Index it landed around 4th (score 57); on the Vals Index it sat 2nd (74.70%), and on Terminal-Bench 2.1 it placed 2nd (80.90%). Across general performance it trailed Claude Fable 5 and OpenAI’s GPT-5.6 Sol. So the honest framing is: elite at coding, top-tier but not the outright leader at everything else.
Verdict: a frontier-class generalist, clearly best-in-class specifically for front-end code.
Efficiency: the quiet win
Moonshot’s 2.5x scaling-efficiency claim over Kimi K2 is harder for outsiders to verify, but the MoE math checks out — activating 1.8% of the network per token is what lets a 2.8T model be served at $3 per million input tokens at all. That price is the benchmark that’ll actually change your bill.
Verdict: plausible and consequential, if you’re paying per token.
What no one else tells you about Kimi K3
Here’s the line the launch coverage keeps burying: “open weights” does not mean “you can run it.” Kimi K3’s weights are roughly 1.4TB in their compressed MXFP4 form, and Moonshot itself recommends a supernode of 64 or more accelerators with fast interconnect. You cannot load this on a single H100, H200 or B200, let alone a 4090 under your desk. The runtime overhead — activations, routing state, and a KV cache stretched across a million-token context — pushes the real memory footprint well past the raw weights. For 99% of people, “open” is a licensing and trust statement, not a practical one. You will rent Kimi K3 from a cloud, exactly like a closed model.
The second thing being glossed over: the license. As of the reveal, Moonshot had not published the full license text, and coverage openly joked about the “license nobody actually read.” If you’re planning to build a product on Kimi K3, that’s not a footnote — the difference between a permissive license and one with commercial or output restrictions can decide whether your business is even legal. Do not assume “open weight” equals “do whatever you want.” Read the actual terms on Moonshot’s release before you commit a roadmap to it. This is the same discipline I urged with Gemini 3.5 Pro: separate the confirmed spec sheet from the launch-day noise.
Kimi K3 vs GPT, Claude and Gemini
Here’s how Kimi K3 sits against the closed frontier and the open field, as of September 2026. Prices and rankings move quickly, so treat this as a snapshot.
| Model | Open weights? | Context | Coding standing | API price (in / out per 1M) |
|---|---|---|---|---|
| Kimi K3 | Yes (~1.4TB) | 1M | #1 frontend at launch | $3 / $15 |
| Claude Fable 5 | No | Large | Top overall; #2 frontend | Premium |
| GPT-5.6 Sol | No | Large | Top overall generalist | Premium |
| Gemini 3.5 Pro | No | Very large | Strong all-rounder | Mid-premium |
vs Claude and GPT: the closed leaders still edge Kimi K3 on all-round intelligence and come with polished tooling, safety layers and support. Kimi K3 counters with open weights, a huge context, and a lower per-token price. vs Gemini: Google wins on ecosystem and multimodal breadth; Kimi wins on openness and raw coding. vs other open models: this is the new open-weight champion by a wide margin — nothing else downloadable is in this tier. If you’re choosing a day-to-day coding assistant rather than a raw model, our Cursor review covers the editor layer that would sit on top of any of these.
Who should use Kimi K3 — and who should skip it
Use it if you’re: a developer who wants elite frontend/coding output at a lower token price; an enterprise or lab that needs to self-host a frontier model for privacy or sovereignty and has the accelerators to do it; a builder who wants to avoid single-vendor lock-in. Best for cost-sensitive coding at scale via API.
Skip it if you’re: a non-technical user who just wants a polished chat assistant (a hosted Claude, GPT or Gemini is smoother); a hobbyist hoping to run it locally (you can’t, realistically); or a company that needs airtight, well-documented licensing today — wait until the terms are unambiguous.
The honest limits
The self-hosting story is aspirational for all but the largest players — 64+ accelerators is a data-center commitment, not a weekend project. The license ambiguity is a genuine business risk until resolved. Its overall-intelligence ranking sits behind the very top closed models. Benchmark leadership this early is preliminary and can shift as others update. And Moonshot has not disclosed the active-parameter count or a knowledge cutoff, so some of the picture is still incomplete. None of that erases the achievement — but “largest open model ever” is a headline, not a purchase decision.
The verdict
Kimi K3 is the most capable open-weight model ever released and a real coding standout — but for nearly everyone it’s an API you rent, not weights you run, so judge it on price and licensing, not the download link.
Frequently asked questions
What is Kimi K3?
Kimi K3 is Moonshot AI’s open-weight large language model with 2.8 trillion parameters, released July 27, 2026. It’s the largest open-weight model to date and ranked first in frontend coding benchmarks at launch.
How many parameters does Kimi K3 have?
2.8 trillion total parameters in a Mixture-of-Experts design with 896 experts, of which only about 16 are active per token — roughly 1.8% of the network at any moment.
What is Kimi K3’s context window?
One million tokens (precisely 1,048,576), with native vision support for images alongside text.
Can I run Kimi K3 on my own computer?
Realistically, no. The weights are around 1.4TB and Moonshot recommends 64 or more accelerators with fast interconnect. It cannot run on a single H100, H200, B200, or any consumer GPU.
How much does the Kimi K3 API cost?
As of September 2026, roughly $3 per million input tokens (cache miss), $0.30 per million cache-hit input tokens, and $15 per million output tokens.
Is Kimi K3 better than GPT and Claude?
At frontend coding it ranked first at launch, ahead of Claude Fable 5. On overall intelligence it’s top-tier but trails the leading closed models like Claude Fable 5 and GPT-5.6 Sol.
Is Kimi K3 really open source?
It’s open-weight, meaning the model weights are downloadable, but the full license terms were not published at reveal. Open weight is not the same as a permissive open-source license, so verify the terms before commercial use.
What makes Kimi K3 efficient despite its size?
Its Mixture-of-Experts design activates only about 16 of 896 experts per token, plus architectural changes Moonshot calls Kimi Delta Attention and Attention Residuals, which it credits for roughly a 2.5x scaling-efficiency gain over Kimi K2.
When was Kimi K3 released?
Moonshot AI released the full Kimi K3 weights on July 27, 2026.
Does Kimi K3 support vision?
Yes, it has native vision capability, so it can process images as well as text within its 1-million-token context.
Who made Kimi K3?
Moonshot AI, the Beijing-based lab behind the Kimi assistant, developed and released Kimi K3.
Where can I try Kimi K3?
You can access it through Moonshot’s official Kimi platform and API, and via third-party API routers such as OpenRouter. The weights are hosted on Moonshot’s official release channels.
I cut through AI launch hype to tell you what a model actually costs, what it’s genuinely good at, and whether you can use it. No vendor allegiances — just what the numbers and the fine print say. See our editorial policy.