Kimi K3's 2.8 Trillion Open Weights Just Landed — Only 1.8% of Them Ever Fire

2026-07-28 · Source: Towards AI - Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure · Depth: Advanced, quick

Summary

Moonshot AI released the Kimi K3 model with 2.8 trillion open weights to Hugging Face on July 27, making it the largest open-weight model by a significant margin. Despite its massive parameter count, only 1.8% (specifically 1.786%) of its 896 experts are active for any given token, meaning roughly 50 billion parameters perform work during a forward pass, while the remaining 98.2% remain idle in GPU memory. This high sparsity still translates to a substantial VRAM requirement of 1,488 GB. This figure challenges common assumptions, as an 8x H200 node falls short by 360 GB, and even an 8x B200 node is insufficient by 48 GB, contradicting several self-hosting guides. The release also highlights the distinction between "open weights" and "open source" due to its specific license clause.

Key takeaway

For AI Architects evaluating large MoE models like Kimi K3 for self-hosting, you must independently verify VRAM requirements. Do not rely solely on published guides, as Kimi K3's 1,488 GB VRAM need exceeds even 8x B200 nodes. Your infrastructure planning must account for the full parameter load, not just active experts, and carefully review licensing terms for "open weights" versus "open source" distinctions.

Key insights

Kimi K3's 2.8 trillion open weights demand 1,488 GB VRAM despite 1.8% sparsity, challenging MoE hosting assumptions.

Principles

• MoE sparsity doesn't negate high VRAM.
• "Open weights" differs from "open source."
• Router scores dictate expert activation.

Method

The article describes arithmetic to calculate VRAM cost for sparse MoE models, considering active parameters and total parameters in memory.

In practice

• Verify VRAM needs for sparse models.
• Review MoE model licensing terms.
• Calculate actual memory footprint.

Topics

• Kimi K3
• Mixture-of-Experts
• VRAM Requirements
• Large Language Models
• Open Weights
• GPU Memory

Best for: AI Engineer, NLP Engineer, Machine Learning Engineer, AI Architect, MLOps Engineer

Related on AIssential

• Before Kimi K3 Goes Open: 8 Secrets Every Developer Needs to Know — Kimi K3 is a 2.8T sparse MoE model where all parameters require VRAM despite sparse activation.
• Kimi K3 Is the Biggest Open Source Model Ever. Almost No One Can Run It. — Kimi K3's massive 2.8 trillion parameters make self-hosting impractical despite its open-source nature, highlighting a compute barrier.
• BREAKING: China's Kimi K3 Just Dropped a 2.8 Trillion Parameter Open AI Model — Kimi K3 introduces a 2.8T-parameter open MoE model with a 1M-token context, setting a new scale benchmark.
• Kimi K3 is downloadable. That doesn't mean you can run it. — Kimi K3's high compute demands challenge traditional open source assumptions, signaling a shift in AI model economics.
• MiniMax M3 vs GLM-5.2 vs Kimi K3: which open-weight model should you actually self-host for agentic — Leaderboard benchmarks often mislead real-world self-hosting decisions for agentic coding models.
• not much happened today — Kimi K3's launch signals a new frontier for open-weight models, challenging top closed systems in specific domains.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Towards AI - Medium.