Skip to content
NewsModel release

Moonshot AI ships Kimi K3 open weights: 2.8 trillion parameters, 1.4 TB footprint, on Hugging Face July 27

· by Pondero Newsdesk

The short version

Moonshot AI released the Kimi K3 open weights on Hugging Face on July 27, 2026, making the 2.8-trillion-parameter sparse MoE model the largest openly available frontier model to date. The MXFP4 footprint of roughly 1.4 TB keeps self-hosting in cloud infrastructure territory.

Moonshot AI ships Kimi K3 open weights: 2.8 trillion parameters, 1.4 TB footprint, on Hugging Face July 27

The full open weights for Kimi K3 arrived on Hugging Face on July 27, 2026, putting Moonshot AI's 2.8-trillion-parameter sparse mixture-of-experts model into the hands of any operator with the infrastructure to run it. The practical floor for that infrastructure is steep: in MXFP4 four-bit precision the weights occupy roughly 1.4 terabytes of GPU memory, which puts self-hosting in the domain of cloud clusters and inference providers rather than individual workstations.

What Moonshot built

Kimi K3 carries 2.8 trillion total parameters in a sparse MoE architecture. Only 16 of 896 experts fire per token, so the active parameter count runs to about 50 billion per token, giving the model per-token compute closer to a large dense model than to its headline figure, per Moonshot's announcement blog.

Two new architectural primitives distinguish K3 from prior Kimi releases. Kimi Delta Attention (KDA) changes how information flows across sequence length. Attention Residuals (AttnRes) selectively retrieves representations across model depth rather than accumulating them uniformly at each layer. Moonshot's blog states the combination yields approximately 2.5x better scaling efficiency compared to Kimi K2. The blog also introduces a Stable LatentMoE framework with Quantile Balancing, which derives expert allocation from router-score quantiles and eliminates a sensitive balancing hyperparameter. Each of those efficiency claims comes from Moonshot's internal evaluations and has not been replicated by a third party.

Training used MXFP4 weights with MXFP8 activations. The context window runs to 1 million tokens. Native support covers text, images, and video within the same model. Built-in capabilities include tool calling, browsing, and multi-step planning.

Moonshot's blog includes one detail that illustrates how the model was used during its own development: an early version of Kimi K3 performed the majority of the team's GPU kernel optimization work in the late stages of training.

What 1.4 TB means in practice

MXFP4 four-bit quantization reduces the weights from a full 16-bit footprint of roughly 5.6 terabytes down to about 1.4 terabytes, per TECHi's technical breakdown. That 1.4 TB figure covers the weights only; the KV cache for a 1M-token context window adds further memory on top.

Moonshot recommends running K3 on supernode configurations with 64 or more accelerators. KDA also changes how prefix caching works, so Moonshot contributed a compatible KDA prefill-cache implementation to the vLLM community alongside the weight release to avoid forcing operators to choose between architectural accuracy and inference efficiency.

The license governing the weights was not published before the Hugging Face release date, per TECHi. Commercial self-hosting decisions should wait for Moonshot to publish the terms.

API access has been live since July 16, per TECHi. Kimi API pricing sits at $0.30 per million tokens for cache-hit input, $3.00 per million for cache-miss input, and $15.00 per million output tokens, per Moonshot's product page. Moonshot states the cache hit rate exceeds 90% in coding workloads, making the effective input price closer to the lower $0.30 figure for code-heavy use cases.

What the benchmarks actually show

Moonshot's own announcement leads with an explicit caveat: "While its overall performance still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol, Kimi K3 demonstrated frontier-level performance across our evaluation suite," per the Kimi blog. That is a notably candid framing for a product launch, and it sets a realistic baseline for evaluating the benchmark tables that follow.

On specific kernel optimization tasks run across NVIDIA Hopper GPUs and an alternative GPGPU, K3 "substantially outperformed Opus 4.8, GPT 5.6 Sol" and performed competitively with Claude Fable 5, per Moonshot's evaluation. The company flagged that its Claude Fable 5 numbers came from a third party and may include fallback behavior.

Across the broader benchmark suite, including FrontierSWE, PostTrain Bench, and Terminal-Bench 2.1, K3's scores cluster above Claude Opus 4.8 and GPT 5.5 and below Claude Fable 5 and GPT 5.6 Sol on most measures, per the detailed footnotes in Moonshot's announcement. The evaluations use different harnesses across models, which Moonshot flags explicitly in each footnote; direct numeric comparisons across harnesses carry uncertainty.

Why it matters for AI tool builders

The weight release changes the option set for operators running workloads that require long context, native multimodal reasoning, or agentic loops without sending data to a third-party API. At 1 million tokens, K3 can process full codebases or extended document sets within a single context. The native video support is less common at frontier scale in an openly released model.

The 1.4 TB footprint keeps the door closed for teams without access to large GPU clusters. Community quantization to GGUF or 8-bit formats has followed every major open weight release this year; those formats typically reduce the footprint by 40 to 60 percent, which would bring K3 closer to the range of mid-size GPU clusters, though at some cost to benchmark performance. That quantization work will determine how broadly K3 reaches outside hyperscale infrastructure.

The $3.00/MTok cache-miss API price is accessible for evaluation. For teams with repetitive coding workloads where the cache-hit rate is high, the $0.30/MTok cache-hit price makes extended trials relatively affordable before any infrastructure commitment.

What to watch next

Three things to track over the next two to four weeks: the license terms Moonshot publishes with the weights, which will determine whether commercial self-hosting is viable; the first third-party evaluation outside Moonshot's own harnesses; and whether community-quantized versions of K3 reduce the footprint enough to serve teams with 8-to-16 GPU nodes.

Sources