Top AI Model Releases This Week: NanoChat, Ring‑1T, Qwen3‑VL, Veo 3.1, Claude Haiku 4.5

Karpathy, former Tesla AI director, released the open‑source project nanochat on GitHub (https://github.com/karpathy/nanochat). The repository, under 8,000 lines of code, reproduces the full ChatGPT workflow. Within three days it attracted over 21.7k stars, and users can train a functional “mini‑ChatGPT” on a single GPU in about four hours for roughly $100.

Ant Group’s Ring‑1T Trillion‑Parameter Model

Ant Group open‑sourced Ring‑1T , a trillion‑parameter “thinking” model built on the Ling 2.0 MoE architecture and pre‑trained on 20 TB of text. The model uses Ant’s proprietary reinforcement‑learning system ASystem for inference training, supports a 128 k context window, and in several international competitions and benchmark tests matches or exceeds top closed‑source models. The model is available on Hugging Face (https://huggingface.co/inclusionAI/Ring-1T-preview) and can be tried at https://ling.tbox.cn/chat.

Alibaba’s Qwen3‑VL Visual Language Models

Alibaba’s Tongyi Qwen team announced the 4 B and 8 B variants of Qwen3‑VL , each offering Instruct and Thinking modes. In dozens of authoritative benchmarks these models surpass Gemini 2.5 Flash Lite and GPT‑5 Nano, positioning them among the leading multimodal LLMs. Despite strong results, the 235 B version failed a simple “six‑finger” test, highlighting current limitations in multimodal model maturity. A demo is accessible at https://chat.qwen.ai/.

Google Veo 3.1 Video Generation Model

Google unveiled Veo 3.1 , an upgrade to Veo 3 that can generate 8‑second 720p or 1080p videos with native audio, delivering high‑fidelity visuals and synchronized sound. Key improvements include better prompt adherence, extended scene continuity for minute‑long shots, and audio‑visual narrative synchronization that gives the model director‑level creation and editing capabilities. The release marks a shift toward “directed and edited” AI video generation. More details are on the project blog (https://blog.google/technology/ai/veo-updates-flow/) and demos are available via Gemini (https://gemini.google.com/) and Lovart (https://www.lovart.ai/zh).

Alibaba Qoder CLI AI Coding Assistant

Alibaba introduced Qoder CLI , a command‑line AI programming assistant that integrates top‑tier coding models with a lightweight agent framework. It offers code generation, comprehension, automated engineering tasks, and global code indexing. The tool’s website is https://qoder.com/cli. The author notes that product differentiation now hinges on model choice versus product design, and mentions Alibaba’s earlier CLI tools such as Qwen Code CLI and iFlow CLI.

Anthropic Claude Haiku 4.5

Anthropic released Claude Haiku 4.5 , priced at one‑third of Claude Sonnet 4 and delivering more than double the speed. On the SWE‑bench Verified benchmark, which measures real‑world programming tasks, Haiku 4.5 matches the performance of Sonnet 4 from five months earlier and surpasses it on certain tasks like “Computer use.” The model’s cost‑efficiency aligns with earlier industry observations that smaller LLMs can handle repetitive, single‑purpose tasks faster and cheaper, suggesting a future where large and small models collaborate for optimal cost‑performance balance. The demo is at https://claude.ai/.

Overall, the week underscores a rapid expansion of open‑source and commercial AI models, with a clear trend toward smaller, more cost‑effective LLMs that still achieve competitive benchmark results, while multimodal and video generation capabilities continue to mature.