Kimi K3 Is the World's First Open 3T-Class Model — and It's Almost Here

Moonshot AI has officially unveiled Kimi K3, a 2.8-trillion-parameter model that claims the title of the world's first open 3T-class model. The announcement, covered in the official Kimi K3 blog post, signals a meaningful shift for teams building AI-powered workflows on their own infrastructure: a model that approaches frontier-level reasoning — with open weights arriving by July 27, 2026.

The model is available today on Kimi.com, Kimi Work, Kimi Code, and the Kimi API. But the real event for self-hosted teams is the imminent open-weight release. For the first time, a model in this parameter class will be downloadable and runnable outside a single vendor's cloud.

What Kimi K3 Actually Is

Kimi K3 is built on two architectural innovations: Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), designed to improve how information flows across both sequence length and model depth — critical for handling the model's 1-million-token context window.

The model uses a Mixture of Experts (MoE) design with substantial sparsity: it activates only 16 out of 896 experts per forward pass, paired with what Moonshot calls a Stable LatentMoE framework. This sparse activation strategy is what makes a 2.8T-parameter model practical to run — most parameters sit dormant during any given inference call.

Combined with refined training and data recipes, these structural changes yield what Moonshot describes as an approximately 2.5× improvement in overall scaling efficiency compared to their previous Kimi K2 model — meaning the same compute budget produces significantly more capability.

The model ships with native vision capabilities and defaults to maximum thinking effort at launch, with configurable low- and high-effort modes planned for later updates.

Performance: Close to the Frontier

On Moonshot's evaluation suite, Kimi K3 consistently outperformed all other tested models — with two exceptions. It still trails Claude Fable 5 and GPT 5.6 Sol in overall performance. But the gap is narrow enough that, for many practical workloads, the distinction may not matter — especially once teams can run K3 on their own hardware without per-token API costs.

The full benchmark details are promised alongside the technical report at weight release time. What Moonshot has published now are detailed coding evaluations, and they're the most revealing part of the announcement.

Coding: The Strongest Signal

Kimi K3's most compelling demonstrations come from long-horizon software engineering tasks — exactly the kind of work where autonomous AI agents earn their keep.

Kernel Optimization

Moonshot tested the model's ability to optimize GPU kernels across four tasks spanning AttnRes, KDA, and a 512-head-dimension MLA kernel. Each model worked independently in an identical sandbox with up to 24 hours to profile, rewrite, and benchmark across NVIDIA H200 and GPGPU hardware from an alternative vendor.

Results: Kimi K3 performed competitively with Claude Fable 5 — noting that Fable 5's results may include fallback behavior and were evaluated by a third party. It substantially outperformed Opus 4.8, GPT 5.6 Sol, and GPT 5.5. Notably, during late-stage development, an early version of Kimi K3 handled the majority of the Kimi team's own kernel optimization work — a model eating its own dog food.

From-Scratch Compiler Construction

In a more ambitious test, Kimi K3 built MiniTriton — a compact Triton-like compiler from scratch, including its own tile-level IR layer over MLIR, optimization passes, and a PTX code-generation pipeline. This isn't patching existing code; it's constructing an entire compilation toolchain.

MiniTriton delivers performance on par with or better than Triton and torch.compile on supported roofline benchmarks, and beats Triton on certain workloads. It also sustains end-to-end nanoGPT training with stable convergence, the loss curve closely tracking the reference — validating the full pipeline on a realistic workload.

Visual Development and Chip Design

Kimi K3 combines 3D reasoning, coding, and vision to turn concepts and images into playable interactive experiences, iterating between code and live screenshots in a "vision in the loop" workflow.

Perhaps the most striking demo: in a single 48-hour autonomous run, Kimi K3 designed, optimized, and verified a chip using open-source EDA tools on the Nangate 45nm library. The resulting 4 mm² chip closes timing at 100 MHz and sustains over 8,700 tokens/s decode throughput in simulation, packing 1.46M standard cells, 0.277 MB of SRAM, and an INT4 MAC array with fused dequantization. A chip built by a model, for a model — a concrete demonstration of long-horizon agentic capability.

Research Acceleration

In computational astrophysics, Kimi K3 autonomously reviewed and cross-validated 20+ papers, then implemented the full numerical pipeline to reproduce the I–Love–Q universal relations. Moonshot says the task completed in approximately two hours — work that would typically require one to two weeks by an experienced researcher.

What This Means for Self-Hosted Teams

The significance for teams building on self-hosted AI isn't just "a bigger model." It's three things converging at once.

First, the capability floor is rising fast. Kimi K3 demonstrates that open models can now handle tasks previously reserved for frontier proprietary systems: multi-hour autonomous coding sessions, compiler construction, research paper implementation. The practical gap between "open" and "frontier" is narrowing to a point where, for many business workloads, it no longer justifies the premium of closed APIs.

Second, the economics shift. Running a model on your own infrastructure — even a large one — changes the cost structure. Instead of paying per-token API rates that scale linearly with usage, teams invest in GPU capacity and run inference at marginal cost. For organizations with sustained AI workloads, the math increasingly favors self-hosting. The self-hosted vs. SaaS trade-off is worth running through a calculator for any team spending more than a few hundred dollars monthly on API calls.

Third, the 1-million-token context window matters for agents. AI agents working with large codebases, long documents, or multi-step research chains need substantial context to maintain coherence. A million tokens — with open weights — means self-hosted agent systems can tackle real-world complexity without constant context compression or chunking strategies.

The Practical Reality

There are caveats worth noting. Running a 2.8T-parameter MoE model at reasonable latency demands serious GPU infrastructure — likely multiple high-end cards. The source doesn't specify exact hardware requirements for self-hosted inference. Moonshot states it is "currently working closely with inference partners and open-source maintainers to align technical details and ensure a reliable rollout across the ecosystem" — a signal that deployment won't be trivial on day one.

The model also defaults to maximum thinking effort at launch, with configurable effort modes coming later. For cost-sensitive self-hosted deployments, the ability to dial down thinking effort will matter — burning full compute on a simple formatting task is wasteful.

The full technical report covering architecture, training details, and evaluations has not yet been published but is promised alongside the weight release.

The Bigger Picture

For nine of the past twelve months, Moonshot's models have set the upper bound of open-model sizes. Kimi K3 continues that trajectory, but the real story isn't the parameter count — it's the demonstrated capability. A model that can build compilers from scratch, design functional chips, and compress two weeks of research into two hours is a fundamentally different tool than what was available to self-hosted teams even a year ago.

The open-weight release on July 27 will be the real test. Inference optimization, community tooling, quantization efforts, and framework support will determine how practically accessible K3 becomes for teams without hyperscaler budgets. But the direction is unambiguous: frontier-class AI reasoning is becoming something you own, not something you rent.

Teams building on self-hosted infrastructure today — whether through platforms like OfficeForge or custom stacks — are positioned to benefit most from this shift. When you control your runtime and your model choices, every new open-weight release is a capability upgrade you can adopt on your own timeline, at your own cost, with your data staying exactly where it belongs.