AI News
AI community posts, news and release updates
Latest updates
Elon Musk announced that Grok 4.6 is expected around August 7 with 1.5T parameters, improved SFT and RL; Grok 4.7 with 2.1T parameters will follow weeks later, better in every way but slower to serve.
- 模型发布
- Grok
- xAI
- Elon Musk
An in-depth article analyzing the Kimi K3 model architecture, explaining its design philosophy, key innovations, and comparisons with other models, helping researchers quickly grasp the technical details.
- Moonshot
- architecture
- Kimi K3
- Discussion
Opus 5 is cool, but we still root for Fable 5... though developers aren't liking it much?
Published Jul 28, 2026A post discussing Opus 5 vs Fable 5, saying Opus 5 is cool but still rooting for Fable 5, though hearing developers don't like it much, questioning the hype.
- Fable 5
- Opus 5
- developer sentiment
- News
Elon Musk reveals Grok 4.6 expected around August 7, larger 4.7 weeks later
Published Jul 28, 2026Musk says Grok 4.6 is expected around August 7, followed by larger Grok 4.7 model weeks later.
- Grok
- xAI
- Elon Musk
- News
SOTA on the Hardest AI Memory Benchmark (BEAM, 10M Tokens) with a Smaller Model
Published Jul 28, 2026A new study achieves SOTA on the BEAM benchmark (10M tokens, the hardest AI memory task) using a model smaller than previous SOTA. This shows efficient architecture and training can beat scale.
- benchmark
- memory
- SOTA
- BEAM
- 10M tokens
- smaller model
Details: Grok 4.6 (1.5T params, improved SFT & RL) releases around Aug 7; Grok 4.7 (2.1T params) follows weeks later, better in every way but slightly slower.
- Grok
- xAI
- 模型参数
- Tutorial
Training Agents Part 3: Train local/open-weight agents with reinforcement learning
Published Jul 28, 2026A tutorial on how to train local or open-weight agents with reinforcement learning.
- reinforcement learning
- open-weight
- agent training
This piece explores how AI agents are poised to infiltrate hardware engineering, from design automation to test optimization, arguing this intersection will drive significant transformation.
- AI agents
- future
- hardware engineering
DeepSeek V4 Flash model achieves 32 tokens per second inference on AMD Ryzen AI MAX+ 395, demonstrating strong local inference performance.
- DeepSeek
- 推理性能
- AMD
- V4
Moonshot AI launches the Kimi Global Ambassador Program, seeking influential individuals who have implemented Kimi K3 in products, agents, or workflows to share their experiences and grow the community.
- Kimi
- 社区
- 申请
- 大使