/ A giant open model just got cheap.
Mike Alvarez Joined TheTweaks as an emerging tech , startup and launches journalist. He comes up with those things or products which are not there yet and are hardly possible to come close to. He covers the latest technology, startups and new product launches and releases. He worked 4 years in a startup company in Austin and later he joined TheTweaks and he is proving himself here with his commendable writing performance. He is also the publisher of a small consumer robotics newsletter.
Moonshot AI has introduced Kimi K3, a 2.8 trillion parameter open weight model that is referred to by the manufacturer as the world’s first 3T-class system of its kind. The latest product from Moonshot is characterized by native vision capabilities, a 1 million token context window and, potentially, the most impactful aspect about the model’s unprecedented pricing compared to all major competitors on the frontier market.
As mentioned above, Kimi K3 is a sparse Mixture-of-Experts (MoE) model. Rather than utilizing all of 2.8 trillion parameters to process every input, the architecture of the product routes each query through 16 out of 896 expert sub networks. In other words, this particular design makes the model computationally feasible; full activation at this scale would be too costly to run.
As stated by Moonshot AI, this is the ninth time within the last 12 months when a Kimi release sets a new upper bound for the open model size, making the lab the leader in terms of aggressiveness in this competition. The key competitors are such labs as GLM and DeepSeek.
Two innovations are responsible for the increased performance Kimi Delta Attention (KDA) and Attention Residuals (AttnRes).
KDA is a hybrid linear attention mechanism meant to speed up the processing of a long context. As reported by the company, it improves decoding speeds at the million token range. Meanwhile, AttnRes alters the mechanism through which the information is passed between layers, retrieving relevant representations through depth instead of stacking them homogeneously, thus increasing the training efficiency without increasing the cost.
Combined with a more aggressive MoE routing system called Stable LatentMoE, this innovation package is said to improve the scaling efficiency of the model by 2.5 times compared to the previous Kimi release, Kimi K2.
Moonshot does not pretend Kimi K3 is the most powerful model in the world. Instead, the company states quite clearly that Kimi K3 lags behind the leading proprietary models GPT-5.6 Sol and Claude Fable 5 of Anthropic. However, the gaps are smaller than you may expect from an open model. Additionally, in some areas, Kimi K3 beats both proprietary competitors.
The model excels at long horizon software engineering tasks, autonomous browsing and research tasks and document parsing outperforming its proprietary competitors in some areas. However, Kimi K3 underperforms on frontier reasoning tasks and elite coding tasks here, both proprietary models have a clear advantage.
It is quite different from the majority of open source releases, which generally lag behind proprietary models on all categories.
While the majority of media publications are focusing on the headline 2.8 trillion parameters the more disruptive part of the release is hidden in the pricing table. Specifically, for a cache hit input, Kimi K3 costs $0.30 per million tokens, for a cache miss input $3.00 and for the output $15.00 per million tokens.
Given that Kimi K3 is competitive with the frontier models on multiple real world coding and research tasks, this pricing is rather aggressive. Additionally, Moonshot states that the company’s infrastructure achieves a cache hit rate above 90% in coding tasks, implying that the effective average cost for developers using agentic coding tools can be substantially lower.
Why does it matter? Because enterprises and independent developers choose not the model that is at the very top of the leaderboard but rather the one that has the most efficient cost-to-capability ratio. Given that Kimi K3 delivers 85-90% of the capability of the frontier models for significantly less money, it does not necessarily have to win every benchmark to change how companies build AI powered products.
Moonshot’s release notes contain an unusual list of limitations that the model possesses. It may become unstable if an application does not preserve its “thinking” history between turns or if it switches to K3 during a conversation from another model. Also, Moonshot admits that Kimi K3 is prone to “excessive proactiveness” on long and ambiguous tasks making additional decisions without asking rather than stopping to ask. As a result, the company advises setting explicit behavioral boundaries in system prompts for the production environment. Finally, Moonshot admits a general UX gap with respect to the leading proprietary models.
Currently, Kimi K3 is available across various platforms Kimi.com, Kimi Work desktop app, Kimi Code and the Kimi API, with full model weights to be released by July 27, 2026. The API is compatible with the OpenAI SDK, meaning that it requires only swapping the base URL in most existing integrations.
To sum up, the Kimi K3 is not the smartest model on the market and Moonshot does not claim that. What it means instead is that there is a shift in what “good enough” open source AI can do and how much it costs to perform. From now on, for startups, independent developers and cost conscious enterprises implementing high volume agentic workflows coding assistants, research automation, document processing and more the criterion is not “which model is the smartest,” but rather “which model is smart enough per dollar spent.” On that criteria, Kimi K3 may be considered one of the most disruptive open releases of this year.
Tim Cook’s 15 years as Apple’s chief executive officially ended on September 1 and now the torch has been passed on to John Ternus, who is the company’s long-time hardware. It’s Apple’s first leadership change…
Be respectful and constructive. Have a question or feedback? We’d love to hear from you. Contact us at contact@thetweaks.com