Search Results for "far"

Z-Image is an efficient, open-source image generation foundation model built to make high-quality image synthesis more accessible. With just 6 billion parameters — far fewer than many large-scale models — it uses a novel “single-stream diffusion Transformer” architecture to deliver photorealistic image generation, demonstrating that excellence does not always require extremely large model sizes. The project includes several variants: Z-Image-Turbo, a distilled version optimized for speed and low resource consumption; Z-Image-Base, the full-capacity foundation model; and Z-Image-Edit, fine-tuned for image editing tasks. ...

...Unlike dense models that activate all their parameters for every token, Step 3.5 Flash uses a sparse Mixture-of-Experts (MoE) architecture that selectively engages only about 11 billion of its roughly 196 billion total parameters per token, delivering high-quality reasoning and interaction at far lower compute cost and latency than traditional large models. Its design targets deep reasoning, long-context handling, coding, and real-time responsiveness, making it suitable for building autonomous agents, advanced assistants, and long-chain cognitive workflows without sacrificing performance.

...Because it’s an inference-focused repository, it’s especially useful as a practical reference for running the model, wiring inputs, and producing the autoregressive sequence of gameplay frames. It also serves as a research sandbox for people exploring how far interactive generative models can go with smaller, more accessible checkpoints compared to massive internal systems.

VibeThinker is a compact but high-capability open-source language model released by WeiboAI (Sina AI Lab). It contains about 1.5 billion parameters, far smaller than many “frontier” models, yet it is explicitly optimized for reasoning, mathematics, and code generation tasks rather than general open-domain chat. The innovation lies in its training methodology: the team uses what they call the Spectrum-to-Signal Principle (SSP), where a first stage emphasizes diversity of reasoning paths (the “spectrum” phase) and a second stage uses reinforcement techniques (the “signal” phase) to refine toward correctness and strong reasoning. ...

...Architecturally, it combines Mixture-of-Experts layers with lightning attention, enabling the model to support a native context length of 1 million tokens while using far fewer FLOPs than comparable reasoning models for very long generations. The team emphasizes efficient scaling of test-time compute: at 100K-token generation lengths, M1 reportedly uses only about 25 percent of the FLOPs of some competing models, making extended “think step” traces more feasible. M1 is further trained with large-scale reinforcement learning over diverse tasks.