Dream
One of the representative works on text diffusion language models. It generates text through a denoising process and, on certain reasoning tasks, performs comparably to autoregressive models of the same scale. It serves as a reference point for evaluating whether diffusion-based text generation can match established autoregressive approaches in quality and task performance.
Timeline (last 90 days)
Diffusion language models (such as LLaDA and Dream) use iterative masked decoding, starting from a fully masked sequence and progressively revealing tokens over several denoising steps, with each step applying bidirectional attention over the entire sequence, so the representations of already revealed tokens are updated by newly revealed tokens.
On the LLaDA and Dream models, evaluated on four benchmarks—GSM8K, HumanEval, MATH, and MBPP—dLLM-KVQ achieved the highest average score among all quantization methods in the reported 4-bit and 2-bit prefix/dual cache settings.
Dream has received $17 million in funding from the well-known venture capital firm a16z.
Dream is a startup founded by former Palantir executives, with the goal of transforming the construction industry using AI world models.
Dream's model training relies on more than 100 million square feet of building permit data.
Dream explicitly regards Autodesk as a competitor.
Dream has not yet disclosed product details.
Some analyses believe that "100 million square feet of building permit data" may be the coverage of the training data rather than the exact training set size.
Some hold the view that Dream is more likely to be a supplement to the Autodesk ecosystem rather than a direct replacement.
Dream's Manhunt series videos claim all chase footage is real and unedited, but the DarkSpark development tool found 46 hidden edit points by analyzing subtle changes in the in-game HUD.
5 more timeline events
All Facts (15)
Diffusion language models (such as LLaDA and Dream) use iterative masked decoding, starting from a fully masked sequence and progressively revealing tokens over several denoising steps, with each step applying bidirectional attention over the entire sequence, so the representations of already revealed tokens are updated by newly revealed tokens.
50%UnverifiedOn the LLaDA and Dream models, evaluated on four benchmarks—GSM8K, HumanEval, MATH, and MBPP—dLLM-KVQ achieved the highest average score among all quantization methods in the reported 4-bit and 2-bit prefix/dual cache settings.
50%UnverifiedSome analyses believe that "100 million square feet of building permit data" may be the coverage of the training data rather than the exact training set size.
50%UnverifiedDream's model training relies on more than 100 million square feet of building permit data.
50%UnverifiedDream explicitly regards Autodesk as a competitor.
50%UnverifiedDream has not yet disclosed product details.
50%UnverifiedDream is a startup founded by former Palantir executives, with the goal of transforming the construction industry using AI world models.
50%UnverifiedSome hold the view that Dream is more likely to be a supplement to the Autodesk ecosystem rather than a direct replacement.
50%UnverifiedDream's Manhunt series videos claim all chase footage is real and unedited, but the DarkSpark development tool found 46 hidden edit points by analyzing subtle changes in the in-game HUD.
50%UnverifiedIn normal Minecraft rendering, the hotbar and the inventory GUI are driven by the same render loop, and their state updates occur in the same frame; in Dream's "boat rescue" segment, the hotbar updates one frame faster than the inventory GUI, indicating the video was composited or edited frame by frame.
50%UnverifiedDarkSpark turns the game ticks that the hunger bar could correspond to into a searchable database; combined with computer vision to recognize the hunger bar pattern in a video, the in-game time corresponding to each video frame can be estimated; timestamps in an unedited video should increase monotonically, whereas Dream's videos show multiple cliff-like drops.
50%UnverifiedDream has received $17 million in funding from the well-known venture capital firm a16z.
50%UnverifiedUnsloth now supports training and chatting with text diffusion language models such as LLaDA, Dream, SDAR, LLaDA2, Nemotron-Labs-Diffusion and DiffusionGemma.
50%UnverifiedUnsloth v0.1.906 adds DiffusionTrainer, which can train text diffusion language models such as DiffusionGemma, LLaDA, LLaDA2, Dream, Nemotron and SDAR.
50%UnverifiedMeta and associated labs have introduced the Dream model, which completes image understanding and text-to-image generation with a single model by first generating a draft and then self-evaluating and refining it, with processing speed reportedly improved by approximately 10%.
50%