Kimi K3 Launches: What the 2.8T Open Model Changes
Kimi K3 is the model launch of the moment for a good reason: Moonshot AI has pushed an open model into a scale class previously associated with closed frontier systems. But the parameter count is only the headline. The more practical story is a model designed for long-running coding and knowledge work, with native vision, a 1-million-token context window, and API pricing that rewards repeated context.
Here is the short version: K3 looks most interesting when an agent must keep working across a large codebase, a long research corpus, or a visually grounded task—not when you only need a quick, cheap response.
What Moonshot AI announced
Kimi K3 has 2.8 trillion total parameters and uses a highly sparse mixture-of-experts design. It activates 16 of 896 experts for a token, rather than using every parameter on every step. Moonshot calls it the first open model in the 3-trillion-parameter class.
The architecture combines three notable ideas:
- Kimi Delta Attention (KDA): a hybrid linear-attention design intended to handle long sequences more efficiently.
- Attention Residuals (AttnRes): a mechanism for selectively retrieving representations across model depth.
- Stable LatentMoE: Moonshot's sparse expert framework, activating 16 of 896 experts.
Moonshot says these changes give K3 about 2.5× the overall scaling efficiency of K2. That is a first-party claim, and the detailed technical report is not available yet, so it should be treated as an early architecture result rather than settled independent evidence.
Why K3 is getting so much attention
It is built for long-horizon work
Kimi positions K3 around sustained engineering and knowledge-work sessions. The model is meant to navigate large repositories, coordinate terminal tools, work with research corpora, and continue through many steps with limited supervision.
Its 1M-token context matters here, but context size alone is not the breakthrough. The real test is whether the model can keep a coherent plan, preserve tool state, and recover from mistakes late in a long run. Kimi's launch examples—kernel optimization, a small GPU compiler, chip design, and scientific-computing workflows—are aimed directly at that question.
Vision is inside the coding loop
K3 accepts image and video inputs and is designed to use visual feedback during software work. Kimi highlights frontend engineering, game development, CAD, and other tasks where an agent can inspect a screenshot, change code, and look again.
That is more useful than treating vision as a separate image-description feature. A model that can alternate between terminal state and rendered output can close more of the implementation loop by itself.
The context pricing is unusually simple
The Kimi API charges one flat rate across the full context window:
- $0.30 per 1M cached input tokens
- $3.00 per 1M uncached input tokens
- $15.00 per 1M output tokens
There is no higher price tier for long prompts. Context caching is automatic, and Kimi says its API sees a cache-hit rate above 90% in coding workloads. Repeated repository context could therefore be much cheaper than the sticker input price suggests—if the prefix stays stable.
See the full Kimi K3 pricing page for worked cost examples and limits.
What developers need to know before switching
K3 is not a drop-in upgrade in every agent harness.
First, thinking is always enabled. The API currently accepts only reasoning_effort="max", and reasoning tokens are billed at the output rate. A concise visible response can still use a substantial number of billed tokens.
Second, Kimi says the complete assistant message must be preserved between turns. If an agent discards the model's reasoning history or switches into K3 midway through a session, output quality may become unstable.
Third, several sampling controls are fixed. Kimi's quickstart tells developers to omit temperature, top-p, penalties, and n rather than tune them as they might for another API.
Finally, Kimi says its web-search tool is being updated and is not currently recommended for production. Developers can still connect their own search tool, but they should not assume every first-party capability is equally mature at launch.
Is Kimi K3 actually open source today?
Kimi describes K3 as an open model, but the rollout is staged. The API and Kimi products are available now; Moonshot says the full model weights will be released by July 27, 2026. The technical report and more complete architecture and evaluation details are also still forthcoming.
So the precise description today is: K3 is an announced open-weight model available through hosted products and API, with the actual weight release scheduled shortly after launch. Self-hosting claims should wait until the files, license, serving recipes, and hardware requirements can be inspected.
Those hardware requirements will be significant. Moonshot recommends supernode configurations with 64 or more accelerators for deployment. Open weights expand who can inspect and adapt a model, but they do not make a 2.8T model cheap to run.
The bigger takeaway
K3 strengthens a trend that has become hard to ignore: open models are no longer competing only on benchmark scores or low serving cost. They are being designed as full agent engines—with long context, native multimodality, tool coordination, and hours-long task execution as primary product requirements.
The launch also makes caching a central part of model economics. K3's uncached input price is ten times its cached rate. For agent builders, prompt layout and context reuse are now architecture decisions, not small billing optimizations.
Kimi K3 still needs independent testing, its weights and technical report are pending, and Moonshot itself acknowledges a user-experience gap versus the strongest proprietary models. Even with those caveats, K3 is a serious launch: it moves the open-model conversation from “how close is the benchmark?” toward “how much of a real project can the model finish?”