News reporting focused on AI and machine learning, covering the companies behind these technologies, their real-world applications, and the ethical concerns they raise. This includes areas like generative AI (large language models, text-to-image and video), speech tech, and predictive analytics.
ByteDance, the parent company of TikTok, is reportedly training an AI model with up to 10 trillion parameters, which would make it the largest Chinese AI model ever built. The model, currently in pretraining, is expected to rival Anthropic's top system in scale. ByteDance has avoided distillation for over a year, focusing on original training data.
Alibaba released Qwen3.8-Max, its largest and most capable AI model, claiming it rivals top US systems from Anthropic and OpenAI. The model, with 2.4 trillion parameters, ranks near the top on Arena.AI, and Alibaba will release its weights next week, marking a return to open-weight strategy amid intensifying US-China AI competition.
Cornell Tech researchers have developed an optical receiver that directly alters memory using photocurrents from a projected light matrix. The technology could update AI model parameters in robots and data centers with lower energy consumption than traditional electrical links, though commercial use remains distant.
Poolside has released Laguna S 2.1, its third coding model in three months. The mixture-of-experts model uses 8 billion active parameters and focuses on better behavior during long agentic sessions. It outperforms other agentic coding models in its weight class and sometimes approaches systems 10 to 20 times its size.
Kimi has released K3, a multimodal open-weight model with 2.8 trillion parameters and a one million token context window. In benchmarks, it approaches Claude Fable 5 and GPT-5.6 Sol but trails them slightly. Priced at $3 per million input tokens and $15 per million output tokens, K3 signals an end to ultra-cheap Chinese AI models, though it remains cheaper than top Western competitors.
Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, has released Inkling, an open-weights multimodal model with 975 billion parameters. The model leads U.S. open-source models on the Artificial Analysis Intelligence Index but trails top Chinese models. It shows strong agentic performance but has a 63 percent hallucination rate.
Xiaomi launched MiMo-V2.5-Pro, an open-weight mixture-of-experts model with 1.02 trillion parameters. It completed a full compiler project in 4.3 hours during internal tests and performs close to Claude Opus 4.6 on coding benchmarks. The model uses 40 to 60 percent fewer tokens than competitors like Claude Opus 4.6 and Gemini 3.1 Pro.
IBM's Granite Team details the construction of Granite 4.1, a set of decoder-only LLMs in 3B, 8B, and 30B sizes trained on about 15 trillion tokens through five pre-training phases. The models receive supervised fine-tuning on 4.1 million curated samples and multi-stage reinforcement learning. The 8B instruct version equals or beats the prior Granite 4.0-H-Small 32B-A9B MoE model using a basic dense setup with fewer parameters. All models appear under Apache 2.0 license.
DeepSeek released preview versions of DeepSeek V4 Flash and V4 Pro, mixture-of-experts models with 1 million token contexts. The Pro version features 1.6 trillion parameters, the largest open-weight model. They match top models on reasoning and coding benchmarks while offering lower prices.
Alibaba released Qwen3.6-35B-A3B, an open AI model that tops Google's Gemma 4-31B across agentic coding benchmarks like SWE-bench Verified at 73.4 versus 52.0. The mixture-of-experts model uses only three of its 35 billion parameters at once to reduce compute costs. It also performs well on reasoning tests and matches Claude Sonnet 4.5 in some multimodal areas.