Moonshot AI released its flagship Kimi K3 model on July 16, a 2.8-trillion-parameter Mixture-of-Experts system that AI researcher and Interconnects author Nathan Lambert describes as a genuine watershed moment for open-weight AI. The full model weights are scheduled for release on July 27, a promise the surrounding commentary treats as the operative assumption shaping how significant the release actually is.
Architecturally, K3 introduces two updates Moonshot calls Kimi Delta Attention and Attention Residuals, designed to improve how information moves across long sequences and through the model's depth. Kimi Delta Attention replaces standard quadratic self-attention in most layers with a linear hybrid, which cuts key-value cache memory needs and speeds up decoding at context lengths up to one million tokens. Attention Residuals let each transformer layer selectively pull representations from any earlier layer rather than only accumulating them uniformly, which Moonshot's own measurements put at roughly 25 percent higher training efficiency for under 2 percent additional compute cost, figures that have not been independently verified.
On independent evaluation boards, K3 has already placed first on one of LMArena's rankings and landed close to Claude Opus 4.8 on Artificial Analysis's Intelligence Index, at 57.11, putting it roughly level with GPT-5.5 though still behind Claude Fable 5 and GPT-5.6 Sol. Lambert's central argument is less about any single benchmark and more about the trend line: the gap between open and closed models, and separately between American and Chinese labs, has narrowed from a widely cited six-to-nine-month lag to something closer to three to five months.
The release has not been without friction. Moonshot briefly paused new subscriptions on July 20 after demand pushed serving capacity close to its limits, according to reporting from several outlets tracking the launch. Running the full open weights independently is itself a substantial undertaking: at full precision the model reportedly requires around 1.5 terabytes of GPU memory, or roughly 600 gigabytes in INT4 quantized form, with Moonshot recommending deployment across supernodes of at least 64 accelerators to keep routing traffic within a single high-bandwidth interconnect domain.
The timing has also fed directly into an active US policy debate. Axios has reported that the Trump administration is weighing restrictions on access to advanced Chinese AI models operating on US soil, a step that would sit alongside an earlier June executive order gating access to closed American frontier models. Lambert's own position, laid out in the piece, is that heavy-handed restrictions on open-weight models risk lulling policymakers into thinking the underlying capability trend has been addressed, when in his view open models will continue to cross key capability thresholds regardless of what regulation says, given how quickly weights propagate once released. The same weekend K3 launched, Alibaba announced a forthcoming 2.4-trillion-parameter Qwen 3.8 model, also planned as open weight, reinforcing Lambert's broader point that Chinese labs appear to be deepening rather than retreating from an open-release strategy.