AI News
AI community posts, news and release updates
Latest updates
- News
Kimi K3: Open-weight MoE model with 2.8T total parameters, matching frontier models - Technical report breakdown
Published Jul 27, 2026Kimi K3 is a 2.8T total/104B active MoE model achieving comparable performance to frontier models with 2.5x training efficiency over K2. The report covers KDA, Attention Residuals, Stable LatentMoE, and multi-stage post-training.
- MoE
- オープンウェイト
- Kimi K3
- Attention Residuals
- Stable LatentMoE
- 技術報告
- News
Anthropic CEO responds to Jensen's open letter: Never called for banning open weights, but supports safety testing and anti-distillation.
Published Jul 27, 2026Anthropic CEO Dario Amodei replies to Nvidia CEO Jensen Huang's letter, stating they never advocated a ban on open-weight models, but support measures like chip export controls, anti-smuggling, and mandatory safety testing, while disagreeing that open weights always help defenders more than attackers, especially regarding biology risks.
- Anthropic
- AI safety
- Nvidia
- regulation
- open weights
- Dario
- Jensen
Venice is offering a free DIEM-powered API key for Kimi K3 with $1000 daily credit, open to anyone for the next week.
- Venice
- free API
- Kimi K3
- Discussion
Inference engineer finds tokenization time non-negligible for long contexts, fixes it for Kimi K3
Published Jul 27, 2026An inference engineer notes that tokenization time, usually negligible, becomes significant for 100K-1M token inputs with prefix cache hits. Michael fixed it for Kimi K3.
- tokenization
- Optimization
- inference
- Kimi K3
- Discussion
Dario's new essay is full of BS: China being ≤2 years behind is the nightmare
Published Jul 27, 2026Dario opens his latest essay with bullshitting, claiming that China being merely two years behind is the nightmare scenario. The post criticizes this stance and references his other pieces on export controls and AI exponential.
- Anthropic
- AI policy
- China
- export controls
- Discussion
Anthropic's hypocrisy? Claims right to train for free on all world output, but calls competitor training on its output IP theft.
Published Jul 27, 2026A user calls out Anthropic for maintaining it can train for free on all world output, even if authors object, while considering it IP theft if a competitor trains on Anthropic's output after paying for it.
- Anthropic
- copyright
- IP
- hypocrisy
- training data
Shows how to run Kimi K3 on AMD MI350X GPUs using SGLang.
- SGLang
- AMD
- Kimi K3
- MI350X
Amid speculation, we've published our complete views on open-weight models.
- AI policy
- open-weights models
- position
Kimi-K3 from Moonshot AI has been quantized to 1-bit, resulting in a 246GB model. The author claims it's the world's fastest.
- Kimi K3
- 1-bit quantization
- 模型量化
Despite falling oil prices, semiconductor stocks plunged on fears that Chinese companies' development of immersion DUV lithography could lead to oversupply, and that open-source AI model Kimi K3 might reduce demand for advanced AI chips.
- AI
- 中国
- 半導体
- 株価