AI Chat-Group Daily (群聊日报)Jul 29, 2026, 14:25 UTC+8

Kimi K3 fully open-sourced, Jensen's alliance announced, Anthropic left out

2026-07-28 群聊日报

Kimi K3 released not just weights but a 47-page tech report and three core infra components: MoonEP, FlashKDA, and AgentENV. The model has 2.8T total params, 104B activated, 896 experts with 16 selected per token, native vision, and 1M context. Someone ran the full model on 80 RTX 5090s over 25GbE with no HBM, hitting 20 tok/s—the first open-weight frontier model deployed without HBM. Jensen Huang's second tweet announced the Open Safe AI Alliance with 37 founding members including Nvidia, Microsoft, and Hugging Face. Anthropic is the only holdout. Separately, GPT Pro is widely being silently downgraded to mini; Fable's safeguards keep misfiring on harmless system design chats.

Why it matters: Moonshot AI's full open-sourcing of K3 — weights, 47-page tech report, and three training infra components — is one of the biggest domestic AI stories this week. The 2.8T-param MoE architecture is fully documented, and the community has already verified deployment on 80×5090 G...

Read the original ↗Export Markdown

Take

This one's worth opening because K3 went beyond weights. A 47-page tech report plus MoonEP, FlashKDA, and AgentENV—that's the full training and deployment playbook for a massive MoE model. 2.8T total params, 104B activated, 896 experts with 16 picked per token, native vision, 1M context. Spec-wise it's in GPT-5 territory.

The wild part: someone ran the full model on 80 RTX 5090s that same night. 25GbE Ethernet, GDDR7, no NVLink, no HBM—20 tok/s single stream. First open-weight frontier model deployed without HBM, on consumer GPUs. Same cluster pushed GLM-5.2 from 30 to 110 tok/s last week, so there's headroom.

Don't read this as "anyone can self-host K3." An 8× GB300 setup costs $100K and only hits 200–300 tok/s—barely enough for a few users. The license also has a hook: model services making over $20M/year need a separate agreement, so third-party serving is probably dead.

Same day, Jensen Huang posted his second-ever tweet announcing the Open Safe AI Alliance with 37 founding members including Nvidia, Microsoft, and Hugging Face. Anthropic is the sole holdout—neither signed the letter nor joined the alliance. The group chat's patience with Anthropic is running thin. Andrew Ng's take was cleaner: you can stay closed, but you can't stop others from opening up.

Source: AI Chat-Group Daily (群聊日报) · louyu2015.github.io
Published: Jul 29, 2026, 14:25 UTC+8 · Jul 29, 2026