Total 1,098
OpenAI's Codex, launched in May 2025 as a research preview, put a software-engineering agent inside ChatGPT: powered by codex-1 and trained with reinforcement learning, it ran each coding task in its own cloud sandbox.
On 14 May 2025, Google DeepMind announced AlphaEvolve, a Gemini-powered coding agent that designs algorithms; it reported a better compute-scheduling heuristic, a faster 4×4 matrix multiplication method, and gains on open mathematical problems.
On 13 May 2025, the US Commerce Department moved to rescind the AI Diffusion Rule two days before its compliance date, told officials not to enforce it, and issued guidance on Huawei Ascend chips and Chinese AI model training.
On 8 May 2025 ByteDance Seed released Seed-Coder, a family of 8B-parameter open code models under the MIT license, with Base, Instruct and Reasoning variants.
On 7 May 2025 Mistral AI released Mistral Medium 3, priced at $0.40 per million input tokens, and claimed at least 90% of Claude Sonnet 3.7's benchmark performance at about 8x lower cost.
On 1 May 2025, TII released Falcon-H1, the first Falcon family to combine attention layers with state space models in parallel, shipping six sizes from 0.5B to 34B parameters at once.
At the end of May 2025, Kuaishou released the Kling 2.1 video generation series in three quality tiers, with reported improvements to motion, physics simulation and prompt responsiveness.
Xiaomi released the MiMo-7B series in late April 2025, including a base model and an RL-tuned chat model trained with verifiable math and coding problems; Xiaomi reported competitive reasoning results.
DeepSeek released DeepSeek-Prover-V2, an open-source model for theorem proving in Lean 4, with its paper posted to arXiv on 30 April 2025. It solved 49 of the 658 PutnamBench problems.
On 30 April 2025 AWS made Amazon Nova Premier generally available in Amazon Bedrock. The model was positioned as a teacher for distilling smaller Nova models and launched in three US regions.
On 30 April 2025 Microsoft released Phi-4-reasoning and Phi-4-reasoning-plus, 14-billion-parameter reasoning models fine-tuned from Phi-4 and published under the MIT licence.
On 29 April 2025 Alibaba released Qwen3, a model family with hybrid thinking modes, two mixture-of-experts models and six dense models, under the Apache 2.0 license.
On 29 April 2025 Meta launched a standalone Meta AI app for iOS and Android, built on Llama 4. It added a feed of shared prompts, voice conversations and personalised answers based on profile information.
On 27 April 2025, trucks running on Aurora's self-driving system began regular driverless commercial hauls for Uber Freight and Hirschbach between Dallas and Houston, with no one in the cab and no police escort.
On 25 April 2025 Baidu released ERNIE X1 Turbo, a reasoning model that Baidu said outperforms DeepSeek-R1 and DeepSeek-V3, priced at a quarter of DeepSeek-R1.
On 25 April 2025 Baidu launched ERNIE 4.5 Turbo, priced 80% below ERNIE 4.5, and announced full support for MCP, a protocol for connecting models to external tools.
On 25 April 2025 StepFun released Step1X-Edit's inference code, weights and technical report, along with GEdit-Bench, an image-editing benchmark, under Apache 2.0.
On 22 April 2025 Huawei released Qiankun ADS 4, HarmonySpace 5 and the XMC vehicle control system in Shanghai, claiming 600 million kilometres of highway L3 simulation and a 2025 target for L3 commercial capability.
Physical Intelligence released π0.5 on 22 April 2025, a vision-language-action model built on its earlier π0 that the company says can carry out household tasks in homes it never saw during training.
In April 2025 twenty humanoid robot teams raced a half marathon in Beijing on the same 21.0975 km course used by human runners. Tiangong Ultra won in 2 hours 40 minutes 42 seconds.
On 16 April 2025 OpenAI released o3 and o4-mini, billed as its smartest models to date, alongside the open-source Codex CLI; o3 scored 71.7% on SWE-bench Verified and 2727 Elo on Codeforces.
