Kimi K3 Explained: The Largest Open-Source AI Model Ever, Benchmarks and Pricing

On July 17, 2026, Beijing-based Moonshot AI released Kimi K3, a 2.8-trillion-parameter model that instantly became the largest open-weight AI model ever published. The release landed one week before the World Artificial Intelligence Conference in Shanghai and triggered market reactions that many analysts compared to the DeepSeek shock of early 2025. The headline is not just the size. Kimi K3 ships with a 1-million-token context window, native vision, an always-on reasoning mode, and benchmark results that place it third in the Artificial Analysis Intelligence Index, right behind Claude Fable 5 and GPT-5.6 Sol, while beating both on several coding and writing benchmarks. The full model weights are scheduled for release on July 27 under a modified MIT license. In this article we break down what Kimi K3 actually is, the architecture behind it, the published benchmarks, the pricing, and what an open model of this caliber means if you build software or AI agents in production.

What Is Kimi K3?

Kimi K3 is the new flagship model from Moonshot AI, the Alibaba-backed startup behind the Kimi assistant and the earlier K2 family. It is a sparse mixture-of-experts (MoE) system, of its 2.8 trillion total parameters, each token is routed through only 16 of 896 expert subnetworks, roughly 1.8% of the model at a time. That sparsity is what makes a 3T-class model economically viable to serve.

• The largest open-weight model ever, at 2.8 trillion parameters, Kimi K3 is the first open 3T-class model. Weights are scheduled to be published on July 27, 2026 under a modified MIT license, with a technical report to follow.
• 1M-token context, multimodal by default, kimi K3 processes text, images and video in a unified architecture, with a maximum context window of one million tokens.
• Always-on reasoning. The model runs with maximum thinking effort by default. Moonshot has announced lower-effort modes for latency-sensitive workloads.
• Available today, weights soon, the model is live on kimi.com, Kimi Code and the Kimi API since launch day; the free tier has been overloaded since release.

The Architecture, how Moonshot Scaled to 2.8T

Moonshot attributes the jump to a set of in-house architectural innovations that together deliver roughly 2.5 times the scaling efficiency of Kimi K2:

• Kimi Delta Attention (KDA), a hybrid linear attention mechanism that speeds up decoding for long sequences, up to 6.3x faster at million-token contexts according to Moonshot.
• Attention Residuals (AttnRes). A drop-in replacement for residual connections that selectively retrieves representations across depth instead of accumulating them uniformly, which the company credits with consistent scaling gains.
• Stable LatentMoE with Quantile Balancing. The routing framework that activates 16 of 896 experts per token. Expert allocation is derived directly from router-score quantiles, removing a notoriously sensitive balancing hyperparameter.
• Per-Head Muon and quantized serving, the optimizer treats attention heads independently for more adaptive learning at scale, and the model ships with MXFP4 weights and MXFP8 activations for broad hardware compatibility.

Benchmarks, third Overall, First in Several Categories

The published numbers are remarkable for an open model. Kimi K3 ranks third in the Artificial Analysis Intelligence Index with 57 points, behind Claude Fable 5 (60) and GPT-5.6 Sol (59), and it wins outright on several published benchmarks:

Two caveats before drawing conclusions. First, most of these numbers come from Moonshot's own launch material, and the independent replication cycle has barely started. Second, Moonshot itself acknowledges a remaining gap versus Claude Fable 5 and GPT-5.6 Sol on overall capability, and notes known issues such as excessive proactiveness during task execution and sensitivity to preserved thinking history. The picture will sharpen once the weights and the technical report land on July 27.

• Artificial Analysis Index, kimi K3 57, Claude Fable 5 60, GPT-5.6 Sol 59
• FrontierSWE (Dominance), kimi K3 47.4, Claude Fable 5 34.2, GPT-5.6 Sol 37.6
• PostTrain Bench, kimi K3 92.4, Claude Fable 5 90.1, GPT-5.6 Sol 88.3
• Arena AI Frontend Code (Elo), kimi K3 1,679, Claude Fable 5 1,631, GPT-5.6 Sol n/a
• GDPval-AA v2 (Elo), kimi K3 1,687, Claude Fable 5 1,815 (Max), GPT-5.6 Sol 1,747.8 (Max)

Pricing, frontier Capability at a Third of the Cost

The economics are where Kimi K3 puts real pressure on closed vendors. On the Kimi API, list pricing sits at a fraction of what comparable frontier models charge:

• $3 input / $15 output per million tokens, compared with $10 / $50 for Claude Fable 5, that is less than a third of the list price for a model in the same benchmark neighborhood.
• $0.30 per million cached input tokens, cache-hit input is billed at a 90% discount, and Moonshot reports cache hit rates above 90% on coding workloads, where repeated context dominates.
• Cheaper per task, not just per token. Independent estimates put Kimi K3 at roughly $0.94 per completed agentic task, versus $1.04 for GPT-5.6 Sol and $1.80 for Claude Opus 4.8.
• Self-hosting on the horizon, once the weights are public, teams with the hardware can serve K3 themselves, and inference providers will compete on price for the same model, which historically pushes token costs down fast.

What Self-Hosting Kimi K3 Actually Costs

Open weights sound like freedom until you price the hardware. Kimi K3 ships in MXFP4, which compresses the 2.8T parameters to roughly 1.4 TB of weights, and that is before the KV cache that a 1M-token context demands. No single machine on earth runs this, you need a multi-node GPU cluster.

The honest conclusion, for the vast majority of companies, self-hosting K3 makes no financial sense, and the API remains the rational choice. Self-hosting pays off in three scenarios, massive sustained volume where the cluster amortizes, hard data-residency or compliance requirements that rule hosted APIs out, and platforms that spread one cluster across many products. For everyone else, the real value of the open weights is indirect, third-party inference providers will compete to serve the exact same model, and that competition pushes token prices down for you without ever touching a GPU.

• The hardware floor, around 1.4 TB of VRAM, a realistic minimum serving setup is around 16 NVIDIA B200 GPUs (192 GB each, two 8-GPU nodes) or 24 or more H200s (141 GB each), just to load the weights with room for the KV cache and batching. Below that, you are not serving traffic, you are running a demo.
• Renting, roughly $60,000 to $100,000 per month, b200 cloud rental averages about $6.25 per GPU-hour in mid-2026 (ranging from $3.75 to over $14 depending on provider). Sixteen B200s running 24/7 is $43,000 to $72,000 per month at the cheap end of the range, and that is before storage, networking, egress and the engineers who keep vLLM or SGLang alive at 3 a.m.
• Buying, a seven-figure project, a B200 costs roughly $40,000 to $50,000 per unit, so 16 of them is $640,000 to $800,000 in GPUs alone. Add servers, InfiniBand networking, storage and redundancy and the cluster lands comfortably above $1 million in capex, drawing around 30 kW of power, plus an MLOps team to operate it.
• What that same money buys on APIs, say $65,000 per month. On the Kimi API ($15 per million output tokens) that is over 4 billion output tokens per month; on Claude Fable 5 ($50 per million) about 1.3 billion. Very few products burn tokens at a rate that makes owning the metal cheaper than paying per token.
• Kimi K3, aPI price $3 / $15 per 1M tokens. Self-hosting option, yes, weights July 27, around $60K-100K/month cluster.
• Claude Fable 5, aPI price $10 / $50 per 1M tokens. Self-hosting option, no, API only.
• GPT-5.6 Sol, aPI price $5 / $30 per 1M tokens. Self-hosting option, no, API only.
• Claude Opus 4.8, aPI price $5 / $25 per 1M tokens. Self-hosting option, no, API only.

The Market Reaction and the Timeline

The launch sequence matters for understanding the noise. On July 16, TechCrunch reported that Moonshot's upcoming model was expected to close the gap with Claude Opus 4.8. On July 17, Kimi K3 launched on kimi.com and the API, and US tech stocks wobbled in a reaction that Fortune and Yahoo Finance described as DeepSeek flashbacks. On July 27, the full weights are scheduled to be published, days before the World Artificial Intelligence Conference in Shanghai.

The pattern echoes January 2025, a Chinese lab releasing near-frontier capability at radically lower prices, with open weights as the distribution strategy. The difference is scale. DeepSeek R1 was a 671B model; Kimi K3 is 2.8T, and it arrives with vision, a 1M context and an agentic focus from day one.

What Kimi K3 Changes in Practice

Beyond the headlines, three things stand out for teams that ship software on top of these models:

• The open-closed gap is now measured in weeks, not years. An open-weight model sitting one or two points behind the closed frontier resets every make-or-buy conversation. If your workload tolerates a small capability gap, the cost difference is now enormous.
• Price pressure lands on everyone. A frontier-class model at $3 / $15 forces closed vendors to justify their premium. For high-volume products, routing cheap traffic to a model like K3 and reserving closed flagships for the hardest requests is becoming the default architecture.
• Self-hosting frontier AI becomes a real option, with published weights and a modified MIT license, regulated industries and data-sensitive products can run near-frontier AI inside their own perimeter. The bottleneck shifts from access to infrastructure and MLOps capability.

How We Evaluate a Release Like Kimi K3 at Alher Tech

Every model launch arrives wrapped in benchmark charts, and launch weeks are the worst moment to make architecture decisions. We ship AI agents and automations into production, so our job is to separate what a release actually changes from what it merely announces.

If you are wondering whether Kimi K3 belongs in your stack, we can run that evaluation on your real use cases and design the architecture around the answer.

• Evals on your real tasks first. We benchmark Kimi K3 head-to-head against your current models on your actual prompts and workflows. Launch-day benchmarks rarely predict production behavior, especially for a model with acknowledged proactiveness quirks.
• Routing before replacement. The sensible first step is rarely a migration. We add candidates like K3 to a routing layer, send them the traffic where they win on cost, and keep the incumbent for everything else.
• Open weights need an ops plan, self-hosting a 2.8T MoE is not a weekend project. We help teams decide between hosted APIs, third-party inference providers and self-hosting, based on data constraints and real volume, not hype.
• Compliance is part of the eval, data residency, GDPR and vendor jurisdiction matter as much as benchmarks. An open model you can host in the EU changes that conversation, and we factor it in from the start.

Conclusion

Kimi K3 is the strongest statement the open-weight movement has made to date, a 2.8-trillion-parameter multimodal reasoner within touching distance of the closed frontier, at a third of the price, with weights landing on July 27 under a modified MIT license.

Healthy skepticism still applies. Most numbers are self-reported, the known issues are real, and the independent evaluation cycle is just beginning. But even with those caveats, the strategic picture has shifted, frontier-class AI is no longer exclusively a closed, US-hosted product.

If you want help evaluating Kimi K3 against your current stack, or you are planning an AI agent or automation on top of it, that is exactly what we do for our clients.

Related articles

• Kimi K3 specs, benchmarks and rankings on our AI model comparison
• GPT-5.6 Explained, sol, Terra & Luna Models, Pricing and Benchmarks
• GPT-5.6 vs Claude Fable 5, benchmarks, Pricing and Which One to Use