Total 671 · Clear filters
OpenAI held DevDay 2025 in San Francisco on 6 October 2025. It announced AgentKit, the Apps SDK and general availability for Codex, and disclosed usage figures for ChatGPT and its API.
On 1 October 2025 Thinking Machines Lab announced Tinker, its first product: a managed API for fine-tuning open-weight language models. It uses LoRA to share compute across training runs.
In late October 2025 MiniMax released MiniMax-M2, a 230-billion-parameter mixture-of-experts model built for coding and AI agents. It scored 61 on the Artificial Analysis Intelligence Index, which MiniMax said ranked first among open-source models.
On 30 September 2025 OpenAI released Sora 2, a model that generates video together with synchronised audio, and an invite-only iOS app built around a social feed and cameos. Within days, rights holders complained about copyrighted characters.
DeepSeek released the experimental model DeepSeek-V3.2-Exp on 29 September 2025. It introduced DeepSeek Sparse Attention (DSA), a method aimed at cutting the cost of processing long inputs, and lowered API prices by more than 50 percent.
On 29 September 2025 Anthropic released Claude Sonnet 4.5 at the same price as Sonnet 4, alongside the Claude Agent SDK, the toolkit behind Claude Code offered to other developers.
On 28 September 2025 Tencent released the code, weights and technical report for HunyuanImage 3.0, an open-source mixture-of-experts image generation model.
In September 2025 Moonshot AI added OK Computer to Kimi, an agent mode that turns prompts into multi-page websites, editable slides and other media, and the first agent product mode in Kimi.
On 25 September 2025, Google DeepMind announced Gemini Robotics 1.5, a vision-language-action model that reasons in natural language before acting, and Gemini Robotics-ER 1.5, a higher-level planning model offered to developers in preview.
On 19 September 2025 xAI released Grok 4 Fast, a cheaper reasoning model with a 2-million-token context window. It was made available to all Grok users, including the free tier.
At Huawei Connect 2025 in Shanghai, Huawei published its first three-year Ascend chip roadmap, running to 2028, alongside new Atlas SuperPoD systems and a pledge to open source CANN.
On 18 September 2025 Luma AI launched Ray3, which it called the world's first reasoning video model. Paid Adobe Firefly and Creative Cloud Pro customers got unlimited generations for their first 14 days.
On 15 September 2025 OpenAI released GPT-5-Codex, a version of GPT-5 tuned for agentic software engineering, as the default model for its Codex coding agent, with API access following later that month at GPT-5 prices.
On 10 September 2025, Zoox began offering public rides in a purpose-built robotaxi on and around the Las Vegas Strip, booked through its app and backed by a remote support team.
On 10 September 2025 Stability AI released Stable Audio 2.5, an audio generation model the company describes as the first built for enterprise sound production at scale, trained on fully licensed data and offered through its API, partner platforms and on-premises licenses.
On 9 September 2025 Baidu open-sourced ERNIE-4.5-21B-A3B-Thinking, a mixture-of-experts reasoning model with 21 billion total and 3 billion active parameters, aimed at logic, mathematics, coding and tool use.
On 9 September 2025 Baidu unveiled ERNIE X1.1 at WAVE SUMMIT 2025, claiming gains over ERNIE X1 of 34.8% in factuality and 12.5% in instruction following.
In September 2025 Writer released the Palmyra-mini family — a base model and two reasoning variants of 1.5–1.7 billion parameters — under Apache 2.0, aimed at on-device and speed-sensitive use.
ByteDance Seed posted the UI-TARS-2 technical report on 2 September 2025 and released it on GitHub on 4 September. It was trained with multi-turn reinforcement learning.
On 1 September 2025 Meituan's LongCat team released LongCat-Flash, a model with 560 billion total parameters that activates 18.6 to 31.3 billion per token, ships open weights under the MIT license and is aimed at agentic use.
In September 2025, Li Auto began rolling out its VLA Driver model via OTA 8.0 to all AD Max users, reporting sharp increases in assisted-driving usage and take-over intervals.
