AI evaluation is now written into law, funding and lab release decisions, but the evidence base it rests on is thinner than the paperwork suggests. In August, Frontier Security reported that a model called Kimi K3 passed a UK AI Safety Institute benchmark by cloning the answers off GitHub.
The Information reports that Tesla's Optimus V3 is running into trouble with its robotic hands, production line, and training data. The company now builds a few hundred units a week. XPeng's robot has cleared supply chain factory audits, while Changan has set its mass production target for 2028.
The model card lists 890 bytes of cache per token, about a quarter of what V4-Flash uses, and HBM demand at a quarter of the previous generation.
Released on 10 September 2026: a multimodal Mixture-of-Experts with 552 billion parameters activates 8 billion per input token and 16 billion per output token. The architecture changed, not just the scale.
The Cyberspace Administration of China reports 1,112 generative AI services have completed filing as of August 31, 2026. Meanwhile, the AI Safety Governance Framework 3.0 was released during National Cybersecurity Awareness Week.
In the model card's table, V4.1-Flash scores 90.6 on Terminal-Bench 2.1 and 74.2 on DeepSWE v1.1, matching or beating Opus-5.0, GPT-5.6 Sol, K3 and GLM-5.3.
Reuters and The Information report that DeepSeek has hired CITIC Securities to prepare a STAR Market listing and is pushing a second funding round targeting 50 billion yuan. Annualized revenue is said to have passed 1 billion US dollars.
MiMo-V2.6-Pro has 1.02 trillion parameters, 42 billion active. In an open live training run, it scored 46 points on the Artificial Analysis index and took first place among open models.
At AICC 2026, an IDC report split the compute gap into two layers: the ceiling on capability and the scale of intelligence. By 2030 the shortfall could reach 380.9 billion dollars, and stacking cards alone can no longer cover the different loads of Prefill and Decode.
Xiaomi streamed a reinforcement learning run live, and MiMo-V2.6-Pro took the top open-source spot with 1.02 trillion parameters and 42 billion active. DeepSeek also released a vision experiment under the MIT license.
A new DeepSeek paper signed by Liang Wenfeng describes DSec, an agent training system that produces more than 5,000 sandboxes a second, reaches 3 million in a single day and peaks at 380,000 running at once.
On the Artificial Analysis Intelligence Index, Xiaomi's MiMo-V2.6-Pro and the closed-source Grok 4.7 both score 46. Yet one task costs about 29 times more on Grok 4.7. Cost is becoming a benchmark dimension of its own.
Liu Sheng leads the main Attention operator on DeepSeek V4.1. His article "I Had to Bury My Talent Yesterday" hit No. 1 on Zhihu's trending list within a day. Based on his public writing, we examine how AI coding is reshaping the role of engineers.
Kunlun Tiangong showcased Mureka V8 at a music club concert, claiming the model outperformed Suno V5 in musician tests. The CEO frames this as a new genre, not just another tool.
Long context costs usually hit memory, not weights. DeepSeek V4.1 Flash compresses its KV cache to roughly 890 bytes per token, cutting HBM demand to a quarter of the previous generation and SSD demand to an eighth.
V4 Pro was due to go offline on September 14. On September 11, DeepSeek said it would keep serving API calls, with billing unchanged. The earlier plan had routed all V4 Pro requests to V4.1 Flash.
DeepSeek released V4.1 Flash on September 10 Beijing time. The new MoE model carries 552 billion parameters but activates only 8 billion on input, and it cuts KV Cache demand on HBM to a quarter of what the previous generation needed.
WeChat Pay's AI card now works with DeepSeek Harness and OpenClaw. Once authorized, the card handles the entire flow from recommendation to payment within a conversation. WeChat emphasizes main-account isolation and per-order confirmation.