Kimi K3 Explained: Specs, Local VRAM & Pricing (2026)

Kimi K3 is a 2.8-trillion-parameter open-weight Mixture-of-Experts (MoE) model released by China's Moonshot AI on July 16–17, 2026, and it currently ranks #1 on the frontend-code arena. However, the full open-source weights are not yet available — they are promised by July 27, 2026 (as of July 18, 2026 they are not yet listed on Hugging Face). Running it locally is estimated to require roughly 650GB–1TB even at aggressive quantization, meaning this is not a consumer-hardware model but a multi-GPU-server or 1TB-RAM-server model.

Kimi K3 uses Kimi Delta Attention (KDA), Attention Residuals (AttnRes), and Kimi-Linear, a linear-attention hybrid architecture. Kimi-Linear reduces the KV cache by roughly 75% and boosts decoding throughput at the 1M-token context by roughly 6x. Training used quantization-aware training from the SFT stage with MXFP4 weights and MXFP8 activations, and Moonshot AI claims roughly 2.5x scaling efficiency compared to Kimi K2.

On Arena.ai's Frontend Code Arena, Kimi K3 scored 1,679 points, taking the #1 spot ahead of Claude Fable 5 (1,631) and GPT-5.6 Sol (1,618). It won 6 of 7 evaluated domains, placing second only in the Gaming domain. Kimi K3 is also reported to close the performance gap with Anthropic's Opus 4.8.

The K2 line, running from Kimi K2.5 onward, has held steady around 1 trillion parameters. Kimi K3 breaks from that pattern by scaling total parameters 2.8x while claiming a 2.5x improvement in scaling efficiency. That said, active parameters per token remain undisclosed, so accurately understanding real inference cost and hardware requirements will require the official model card that ships alongside the July 27 weight release.

These figures are estimates scaled from the Kimi K2 family's real-world footprint up to K3's 2.8T scale; the definitive spec table won't be available until the July 27 weight release and its accompanying model card. As a reference point, Kimi K2.7, at 1 trillion total parameters, needed roughly four RTX PRO 6000-class cards for its 339GB 2-bit build — K3 is expected to need roughly triple that memory. A realistic minimum setup is a multi-GPU workstation or a server-class machine with roughly 1TB of RAM (CPU inference would likely run at single-digit tokens per second). A 512GB Mac Studio would not have enough capacity to hold the model. The first realistic local context milestone at weight-drop is expected to be around 128k–256k tokens, with the full 1M-token context reserved for API access or large clusters.

Another large open-weight model that emerged around the same period is Inkling from Thinking Machines, at 975 billion parameters and already downloadable. Unlike Kimi K3, Inkling's weights are already public, so its local hardware requirements can be measured directly rather than estimated. Kimi K3 is considerably larger at 2.8 trillion parameters, but its weights remain unreleased until July 27, so every local-run figure discussed here is currently an estimate.