VibeThinker-3B is a small 3-billion-parameter model that performs exceptionally well on the MathQA benchmark, achieving results comparable to models with around 30 billion parameters. The model's strong performance has sparked discussion about its efficiency and capabilities in mathematical reasoning.
The Minimax M3 (4-bit MLX) was benchmarked on a Mac Studio M3 with 512GB storage. Results show token throughput and latency metrics across different prompt sizes, with peak performance at 269.1 tok/s for 8192-token prompts and 172.8 tok/s for a 65k-token prompt, using 228GB of peak memory.
GLM-5.2 has been ranked first on Design Arena, surpassing the previously available Claude Fable 5. The Claude Fable 5 model is now unavailable, contributing to GLM-5.2's top position.
GLM 5.2's API is now live, with model weights available on Hugging Face under MIT license and supported by Ollama. The model offers two thinking modes—High and Max—with 1M context length, priced at $1.4 per 1M input tokens and $4.4 per 1M output tokens, matching GLM-5.1.
GLM-5.2 is the first open-weights model to achieve 80% accuracy on Terminal-Bench and outperforms all other available open models. It also surpasses Gemini, positioning it as a frontier-level model at a significantly lower cost.
GLM-5.2 has secured the second position in the WebDew Arena benchmarking evaluation. The result reflects its strong performance in natural language understanding and generation tasks compared to other models.