Together AI Ships Moonshot's Kimi K3 the Day Open Weights Drop

Kimi K3, Moonshot AI's 2.8-trillion-parameter open-weight model, is coming to Together AI's serverless inference API on July 27 -- the exact same day Moonshot releases the full model weights to the public. That kind of day-zero availability is a deliberate signal: Together AI wants to be the default inference layer for frontier open-weight models the moment they land, not days or weeks later.

The model itself is a big deal. Kimi K3 is the world's first open-source model in the 3-trillion-parameter class. Moonshot AI, the Beijing-based startup backed by Alibaba, benchmarks it as performing neck-and-neck with the most powerful proprietary systems from Anthropic and OpenAI. Whether that claim fully holds up depends on which benchmark you look at -- but the broad picture is clear: open-weight models just got a lot more competitive.

Under the Hood: Why 2.8T Doesn't Mean What You Think

The headline parameter count is real, but the architecture is what makes it practical to serve. Kimi K3 is a Mixture-of-Experts (MoE) model with roughly 2.8 trillion total parameters, a 1-million-token context window, native vision, and always-on reasoning, aimed at long-horizon coding and agent workloads.

MoE (Mixture-of-Experts) means the model is actually a collection of specialized sub-networks called "experts," and only a small subset activates for any given token. Of 896 experts, only 16 fire for any given token -- roughly 1.8% of the network -- so you get the knowledge of a 2.8-trillion-parameter model at a fraction of the compute per step.

The other architectural trick is KDA (Kimi Delta Attention). K3 runs on KDA, a hybrid linear-attention mechanism paired with attention residuals -- what keeps a million-token context affordable instead of ballooning quadratically. Standard attention scales as O(n²) with sequence length, which makes 1M-token contexts prohibitively expensive. KDA sidesteps that. A 1,048,576-token window holds entire repositories, long design docs, and test histories in a single session -- the raw material for repo-scale reasoning rather than snippet completion.