🤖
AI Tool Alert — July 26, 2026Kimi K3 Open Weights — July 27 DropMoonshot AI announced
Kimi K3 full open weights releasing July 27 on Hugging Face under a
Modified MIT license — the first open-weight model to reach the 3-trillion-parameter class.
The specs:- 2.8T parameters — MoE architecture, 16 active experts per pass
- 1M token context window (1,048,576 tokens)
- GPQA Diamond: 93.5% — competitive with Claude Fable 5 & GPT-5.6 Sol
- MXFP4 quantization — ~18x H100 80GB GPUs or 64+ accelerators
- Permissive license — commercial self-hosting allowed
Why it matters for devs:- Self-host frontier-level reasoning without API rate limits or vendor lock-in
- 1M context window = feed entire repo histories in a single prompt
- MoE architecture = only ~16 experts active per request despite 2.8T total params
- Modified MIT = no copyleft, no revenue caps, fine-tune freely
Concrete use case — internal code review pipeline:Closed API at ~$5/review → $10K/month at 100 reviews/week. Self-hosted Kimi K3 at ~$800/month for 400 reviews/week. Breakeven at ~40-50 PRs/day. Data stays on-prem, no rate limits, fine-tune on your codebase style.
🔗
Kimi Blog |
Northflank Self-Host Guide |
HF Overview