« A Tencent diffusion language model capable of producing multiple tokens in parallel, with inference speeds up to 10 times faster than traditional autoregressive LLMs. Compatible with standard causal attention and native KV cache, optimized via vLLM »

Best AI tools forLLM models
AI models to test and tools to learn how to master them
« An open-source model with 40 billion parameters specializing in code that outperforms Claude Sonnet with an 81% success rate on SWE-Bench. Its architecture, trained on rest evolution, guarantees professional-quality code »
« An open-source model optimized for autonomous coding that rivals Claude Sonnet 4.5 while costing 10 times less. Specialized in multilingual coding, tool usage, and long-horizon planning »
« A lightweight model created by Z.AI that achieves open-source SOTA among models of similar size. Optimized for agentic coding with a “think before you act” mechanism, the model also excels at Chinese writing, translation, and long texts »
« An open-source multimodal model with 1T parameters and a swarm agent orchestrating up to 100 sub-agents in parallel. Powerful for front-end code generation from conversations/videos, autonomous visual debugging, and a 256K context window »
« This 745B-parameter Mixture-of-Experts model aims for complete agentic intelligence with planning, tools, web browsing, and context up to 200K tokens, while being published under an MIT license »
« This large multimodal model combines text, vision, and interface interaction in a single system, enabling it to understand screenshots, videos, and documents. It can also reason in multiple steps and handle over 200 languages »
« This version of Grok brings together four professor-level agents who reason together before responding. They correct each other, break down problems, and rely on real-time web search and coding tools to provide more reliable answers »
« This open-source model combines reasoning, image analysis, and advanced coding into a single tool. Powerful and flexible, it can handle long queries and generate more direct and effective responses. A robust alternative to the best models on the market »
« Alibaba's new leading model codes, tests, and debugs entire projects autonomously using a 1-million-token prompt. It generates web interfaces from screenshots, understands images and documents, and integrates directly with Claude Code via the Anthropic API »
« Take advantage of enhanced audio and visual understanding to create comprehensive multimodal applications, ranging from image analysis to audio interpretation. The open-source E2B and E4B models offer optimal memory capacity and computational efficiency. Perfect for inference on devices with limited resources »
« This open-source 400B-parameter model thinks before responding to handle long and complex prompts. It ranks just behind Claude Opus 4.6 on PinchBench and costs 96% less (Apache 2.0 license) »