Mirasol3B: A Multimodal Autoregressive model for time-aligned and contextual modalities · HackerPapers discussionshttps://hackerpaper.kaipfister.de/p/2311.05698/Collected HN comments and submissions linking this paper.enMon, 02 Nov 2026 11:16:22 GMTMirasol3B: A Multimodal Autoregressive model for time-aligned and contextual modalities — mentioned by msoadhttps://news.ycombinator.com/item?id=40424593urn:hackerpapers:mention:2311.05698:40424593Tue, 21 May 2024 05:44:58 GMT<p>Compared to Mirasol3B[1] this is not supporting audio as a modality. What Google has done with Mirasol3B made the demo of &quot;Astro&quot; in Google I/O possible. They do a little of cheating by converting audio to images(spectrogram) and video to 25 photo frames with some sort of attention system to things that change during those frames. So the tokenizer is basically the same for audio and video and images. I believe Meta is going to this direction with multimodality as well. The new GPT voice mode is probably using the same architecture. What&#x27;s mind boggling is that models perform better at the…</p><p><a href="https://hackerpaper.kaipfister.de/p/2311.05698/">Paper and discussion history</a>. A mention is not an endorsement.</p>cs.CV