Kimi K3 Full Weights Officially Released, Taking the 2.8T MoE from API Access to Self-Deployment

Moonshot AI published Kimi K3’s full weights, model card, and technical report on Hugging Face on July 27, fulfilling its previously announced open-weight plan. When K3 initially launched, it was primarily available through Moonshot’s applications and API. Although the company had disclosed its 2.8-trillion-parameter sparse MoE architecture and plans for a subsequent weight release, outsiders could not directly inspect the model or validate self-hosted performance. This release allows researchers and engineering teams to audit, benchmark, and modify the actual weights, shifting the discussion from endpoint capabilities to infrastructure, licensing, and reproducibility.

K3 is a natively multimodal model with 896 experts. It selects 16 experts per token and includes two shared experts. The model has 2.8T total parameters, with approximately 104B activated during each forward pass. Its 93-layer network comprises 69 Kimi Delta Attention (KDA) layers and 24 Gated MLA layers, while Attention Residuals selectively retrieve representations from different depths. Moonshot says Stable LatentMoE and a new training recipe deliver approximately 2.5 times the scaling efficiency of Kimi K2, though external teams have yet to reproduce this result using the same definition.

The released weights use MXFP4 produced through quantization-aware training, while activations use MXFP8. The model card lists a maximum context length of 1,048,576 tokens. Although the low-precision formats reduce storage and compute requirements, deployment remains highly demanding: Moonshot recommends a supernode with at least 64 accelerators and a large, high-bandwidth communication domain. vLLM, SGLang, and Transformers provide launch instructions, but KDA’s prefix cache requires a specialized implementation. The Transformers example also requires `trust_remote_code=True`; teams should review the remote custom code and the Kimi K3 license terms before running it.

The official technical report lists scores of 88.3 on Terminal-Bench 2.1 and 67.5 on DeepSWE. However, the comparison models use different agent harnesses, including Kimi Code, Claude Code, and Codex, and some tasks were recalibrated on H20 hardware, so the scores should not be treated as strictly like-for-like comparisons. The model also depends on preserving the complete thinking history. If an agent switches to another model mid-task, or if the framework does not return the existing reasoning history, output quality may become noticeably unstable. With the weights now available, the next key questions are whether the community can rerun agent benchmarks under the same harness and whether open-source inference backends can reproduce K3’s throughput and long-context capabilities.