Kimi K3: Specs, Benchmarks, and How to Access It (2026)

Moonshot AI's Kimi K3 is the biggest open-weight release ever announced: 2.8 trillion parameters, a 1 million token context window, and weights scheduled to drop July 27. Unlike most frontier launches, this one already has real third-party benchmark data behind it. Here is what K3 is, what the independent numbers actually show, and what it takes to run it.

TL;DR

• Announced: mid-July 2026, API access first. Open weights scheduled for July 27, 2026
• Size: 2.8 trillion total parameters, sparse Mixture-of-Experts (896 experts, 16 active per token). Moonshot has not published a single active-parameter figure
• Context window: 1 million tokens
• Modalities: text and vision, natively integrated, with always-on reasoning
• Quantization: trained quantization-aware, MXFP4 weights. Full model is still ~1.4 TB
• License: not published yet, expected with the weight release. "Open weights July 27" is confirmed intent; the license terms are not
• Benchmarks: strong third-party results already (details below), including topping coding leaderboards
• API pricing (Moonshot): $3 per million input tokens ($0.30 cached), $15 per million output
• Self-hosting reality: this does not fit on one GPU or one node. Think multi-node cluster, 1.4 TB+ of aggregate GPU memory
• The bigger story: independent testing suggests routing between K3 and a frontier model beats using either alone

What Kimi K3 Is

Kimi K3 is Moonshot AI's successor to the K2 line, and it arrived days before Alibaba previewed Qwen 3.8-Max. The two launches frame the same week and opposite strategies: Qwen 3.8-Max is a closed preview with open weights promised vaguely; K3 has a hard date for its weights, July 27.

Architecturally, K3 is a sparse MoE transformer with native vision and a set of efficiency mechanisms Moonshot introduced with this generation, including a hybrid linear attention design (Kimi Delta Attention) aimed at making the 1M context window affordable to serve. The model was trained quantization-aware from the fine-tuning stage, which is why the released weights ship in MXFP4 without a separate lossy conversion step. Moonshot claims a 2.5x scaling-efficiency improvement over K2.

For production teams, the two headline facts are the hard open-weight date and the size. At 2.8 trillion parameters and ~1.4 TB of weights, "open" does not mean "runnable on your workstation." More on that below.

Kimi K3 Benchmarks

K3 is in a different evidence position than most launch-week models: third parties have already tested it.

On Artificial Analysis, K3 scores 57 on the Intelligence Index, placing it 4th of 189 models tracked. On the Vals Index it ranks 2nd of 38. It took first place on Arena WebDev (preliminary). Coding is the standout: community-reported results include best-overall marks on SWE Marathon and Program Bench, and 88.3 on Terminal-Bench 2.1. In one independent architecture evaluation against Qwen 3.8-Max preview, K3 scored 83 to Qwen's 80.

The most interesting third-party data comes from Fireworks AI, which ran K3 against Claude Fable 5 on roughly 1,000 agentic tasks. Their finding was not a winner-takes-all result. K3 outperformed on security, crypto, and long terminal loops; Fable 5 won on multilingual and web/data visualization tasks. Neither model dominated.

The usual caveat applies in reverse here: these are early third-party numbers, not vendor claims, which makes them more credible but still preliminary. Validate on your own workload before making decisions.

The Routing Finding, and Why It Matters

Buried in that Fireworks evaluation is the most important production insight of the month: per-task routing between K3 and the frontier model scored 93% accuracy, higher than either model alone, at up to 50x lower cost on long agentic loops. Their router sent 72-96% of traffic to K3, with the frontier model as the fallback rather than the default.

Read that again: the open-weight model handles the bulk of the work, the expensive frontier model catches the hard cases, and the combination beats both on quality AND cost.

This is the argument for putting a routing layer in front of your models instead of hard-coding one provider, the same pattern we covered in how to run Qwen 3.7 in production. An AI gateway gives you one OpenAI-compatible API in front of multiple models, so shifting traffic between an open model and a frontier fallback is a routing rule, not a rewrite. As specialization data like this accumulates, single-model deployments look more and more like leaving quality and money on the table.

Kimi K3 Pricing

Moonshot's API pricing for K3: $3 per million input tokens on a cache miss, $0.30 per million on cache hits, and $15 per million output tokens. The cached input price is aggressive, clearly aimed at long-context agent workloads that repeatedly read the same context.

Once the weights land, self-hosting changes the cost equation entirely for high-volume workloads, if you have the hardware to serve it. Which brings us to the catch.

How to Run Kimi K3

Via API today. Moonshot's API is live, and K3's OpenAI-compatible interface means standard client code works against it.

Self-hosted, after July 27. This is where the 2.8T scale gets real. The MXFP4 weights alone are ~1.4 TB, which means no single GPU, and no single 8-GPU node, can hold the model. Realistic serving means a multi-node cluster with 1.4 TB+ of aggregate GPU memory and MoE-aware inference; reported recommendations run to 64+ accelerators for production serving, on the newest hardware generations. vLLM support for K3's attention design is expected alongside the weight release.

In practice, that puts self-hosted K3 in reach of teams that can provision multi-node GPU clusters, not single-pod experiments. For most teams, the realistic K3 architecture in August 2026 is the routed one: API access for K3, open-weight deployment for smaller models like GLM 5.2 or Qwen 3.6 where the hardware math works, and a gateway in front of all of it.

Kimi K3 vs Qwen 3.8-Max

The honest summary: K3 currently has the receipts, published pricing, third-party benchmarks, a weights date, while Qwen 3.8-Max has the bigger launch hype and an unverified "second only to Fable 5" claim. That balance can change the day Alibaba publishes real numbers.

Frequently Asked Questions

When did Kimi K3 release? Moonshot announced K3 with API access in mid-July 2026. Open weights are scheduled for July 27, 2026.

Is Kimi K3 open source? The weights are scheduled to be released July 27, but the license has not been published yet. Until the license terms are out, treat "open source" as unconfirmed; "open weights" is the accurate term.

How big is Kimi K3? 2.8 trillion total parameters in a sparse MoE design with 896 experts, 16 active per token. The MXFP4 weights are roughly 1.4 TB.

Can I run Kimi K3 on my own GPUs? Only at serious scale. The model exceeds any single GPU or single 8-GPU node; realistic self-hosting means a multi-node cluster with 1.4 TB+ of aggregate GPU memory. For most teams the API, or a routed hybrid, is the practical path.

Is Kimi K3 better than Qwen 3.8-Max? In the only head-to-head independent test published so far, K3 edged the Qwen 3.8 preview 83 to 80, and K3 has far more third-party benchmark coverage overall. But Qwen 3.8 has no official benchmarks yet, so a definitive comparison isn't possible.

Is Kimi K3 better than Claude Fable 5? Independent testing by Fireworks AI found specialization, not dominance: K3 won on security, crypto, and long terminal loops; Fable 5 won on multilingual and visualization tasks. Routing between them beat both.

What is the cheapest way to use Kimi K3? Cached API input at $0.30 per million tokens is the headline rate. For sustained high volume after July 27, self-hosting on your own cluster can undercut API costs, if the hardware and ops investment makes sense for your scale.

Bottom Line

Kimi K3 is the most consequential open-weight release of 2026 so far: frontier-class third-party benchmarks, published pricing, a hard date for the weights, and a size that redefines what "open model" means operationally.

The strategic takeaway isn't "switch to K3." It's the routing data: an open model handling 72-96% of traffic with a frontier fallback beat both models on quality and cost. That pattern, open weights for volume, frontier for the hard cases, one gateway in front of both, is where production AI is heading.

July 27 is the date to watch. When the weights and license land, we'll cover what it actually takes to serve this thing. Until then, see how the open-vs-closed decision plays out in GLM 5.2 vs Qwen 3.7 Max and the Qwen 3.8-Max preview.