Kimi K3 Open Weights Land on Hugging Face: The World's Largest Open AI Model
On 27 July 2026, Moonshot AI released the open weights for Kimi K3 — a 2.8 trillion-parameter Mixture-of-Experts model and the largest open-weight AI release in history. The weights — 96 shards totalling approximately 1.56 terabytes — were published on Hugging Face under a custom Kimi K3 licence, making them freely downloadable for research and commercial use subject to the licence terms. The release came ten days after Kimi K3's API launch on 16 July 2026, during which the model jumped from eighteenth to first place on the Frontend Code Arena leaderboard with a score of 1,679, overtaking Anthropic's Claude Fable 5. Moonshot paused new subscriptions within 48 hours of the open-weight release, citing GPU capacity constraints following more than 7.2 million views of the announcement post.
Architecture: 2.8 Trillion Parameters, 16 Active Per Token
Kimi K3 uses a Mixture-of-Experts architecture with 896 expert sub-networks. During inference, only 16 of those 896 experts fire for any given token — yielding approximately 50 billion active parameters per forward pass despite the model's full 2.8 trillion parameter count. This design separates the model's knowledge capacity, determined by total parameter count, from the compute cost of inference, determined by active parameters. At roughly 50 billion active parameters, Kimi K3's inference compute cost is closer to that of a 50 billion-parameter dense model than to a 2.8 trillion-parameter one. The model uses MXFP4 four-bit weight quantisation as its baseline storage format, bringing total weight storage to approximately 1.4 terabytes of fast memory. By total parameter count, Kimi K3 comfortably exceeds DeepSeek V4-Pro at 1.6 trillion parameters, the previous largest open-weight model.
Context Window and Benchmark Performance
Kimi K3 supports a context window of 1,048,576 tokens — a 1-million-token window that allows users to load entire large codebases, extended legal documents, or full multi-session histories into a single query without truncation. The model was built specifically for long-horizon coding and end-to-end knowledge work tasks where keeping the entire working context visible simultaneously reduces the need for repeated retrieval calls or context management overhead. On the Frontend Code Arena leaderboard — which benchmarks models on their ability to generate correct, styled front-end code from visual inputs and specifications — Kimi K3 scored 1,679 within hours of its API launch on 16 July, placing first above all evaluated models including Anthropic's Claude Fable 5.
Pricing on the Kimi K3 API
The Kimi K3 API uses cache-tiered input pricing. Cache-miss input tokens cost $3.00 per million. Cache-hit input tokens — where a previously processed context is reused — cost $0.30 per million: a 90 per cent reduction that materially lowers the cost of multi-turn or repeated-query patterns over fixed large contexts. Output tokens cost $15.00 per million. These rates place Kimi K3 at the premium end of frontier model API pricing, reflecting both its capability tier and the infrastructure cost of running a 1-million-token context window at scale.
Running Kimi K3 On-Premises: The 1.4 Terabyte Challenge
Hosting Kimi K3 locally requires approximately 1.4 terabytes of fast memory in MXFP4 precision — a deployment footprint well beyond what a single GPU node can serve. Self-hosting requires a multi-node cluster of high-memory accelerators, with eight H100 or H200 80GB GPUs representing the approximate lower boundary for feasible weight sharding. For most engineering teams, API access is the appropriate path. The open weights are primarily relevant for research organisations that need to fine-tune or modify the model, and for regulated enterprises in banking, healthcare, and government that have data residency requirements preventing use of an external provider's API for production workloads.
What Kimi K3 Means for Indian Engineering Teams
For Indian software and AI teams, Kimi K3's release opens two practical options not previously available at this capability level. First, the 1-million-token context window makes the model credible for use cases where context length has historically been the binding constraint: full-codebase review and debugging, end-to-end legal document analysis, and vernacular language tasks that require holding extensive reference corpora in context simultaneously. Second, for teams operating under data residency constraints — particularly in government, banking, and healthcare — the open weights offer a route to on-premises frontier-capable inference without dependence on any external API provider. Teams evaluating this option should read the Kimi K3 licence carefully; Moonshot AI is a Chinese company, and the commercial use and derivative work terms will determine whether self-hosted deployment is viable for a given organisation's regulatory context.
The Bottom Line
On 27 July 2026, Moonshot AI released Kimi K3 open weights on Hugging Face: 2.8 trillion parameters across 896 Mixture-of-Experts experts, with 16 active per token — approximately 50 billion active parameters and 1.4 terabytes in MXFP4 quantisation. The model supports a 1-million-token context window and scored 1,679 on the Frontend Code Arena leaderboard at launch, placing first above all evaluated models. Kimi K3 is the largest open-weight model ever released by parameter count, exceeding DeepSeek V4-Pro's 1.6 trillion parameters. API pricing is $3.00 per million cache-miss input tokens, $0.30 on cache hits, and $15.00 per million output tokens. For Indian engineering teams, the most actionable properties are the 1-million-token context window for long-context tasks and the on-premises availability of the weights for organisations with data residency requirements.
Frequently Asked Questions
What is Kimi K3 and who built it?+
Kimi K3 is a 2.8 trillion-parameter Mixture-of-Experts large language model built by Moonshot AI, a Chinese AI startup. It uses 896 expert sub-networks, with 16 active per token during inference — yielding approximately 50 billion active parameters per forward pass. The model supports a 1-million-token context window and was designed for long-horizon coding and knowledge work. Its API launched on 16 July 2026, and its open weights — approximately 1.4 terabytes in MXFP4 quantisation — were released on Hugging Face on 27 July 2026 under a custom Kimi K3 licence. At 2.8 trillion total parameters, Kimi K3 is the largest open-weight AI model ever released, exceeding DeepSeek V4-Pro's 1.6 trillion parameters.
How does Kimi K3's Mixture-of-Experts architecture differ from a standard dense model?+
In a dense language model, every parameter is active for every token, and inference cost scales directly with total parameter count. Kimi K3 uses a Mixture-of-Experts architecture with 896 expert sub-networks, but only 16 experts are selected and activated per token — roughly 50 billion active parameters per forward pass, despite 2.8 trillion total parameters. This means Kimi K3's inference compute cost is comparable to running a 50 billion-parameter dense model rather than a 2.8 trillion-parameter one, while the full 2.8 trillion parameters represent the model's total knowledge capacity. The design allows Moonshot AI to build a model far larger than would be economically feasible to run as a fully dense architecture.
What does Kimi K3's 1-million-token context window enable for developers?+
A 1-million-token context window allows developers to load approximately 750,000 words — roughly a 3,000-page document — into a single model query without truncation. In practice, this enables loading an entire large codebase for end-to-end review or debugging, analysing a full legal document corpus without chunking, running extended multi-session conversations without losing earlier context, and processing long-form data pipelines where intermediate outputs must remain visible alongside subsequent inputs. For Indian software teams working on enterprise codebases, regulatory document analysis, or vernacular language tasks requiring extensive reference material, the 1-million-token window makes Kimi K3 credible for use cases where other frontier models require retrieval-augmented generation architectures to compensate for shorter context limits.
What hardware is required to self-host Kimi K3's open weights?+
Self-hosting Kimi K3 requires approximately 1.4 terabytes of fast memory to hold the full MXFP4-quantised weights — a footprint well beyond any single GPU node. In practice, self-hosted deployment requires a multi-node cluster of high-memory accelerators; eight Nvidia H100 80GB or H200 80GB GPU nodes represent the approximate lower boundary for achievable weight sharding. For teams without large GPU cluster access, the Kimi K3 API is the practical path, with cache-miss input pricing at $3.00 per million tokens. Self-hosting is most relevant for research organisations needing to fine-tune or fork the model, and for regulated enterprises in banking, healthcare, or government with data residency requirements preventing API use with a Chinese provider. Those teams should also review the Kimi K3 licence terms before deployment.
Written by
TechPillow Team
Sharing insights on technology, product development, and the Indian tech ecosystem.