{"rewrite":{"id":"r_6a6eac4ed593aa1de2ee86d6","clusterId":"c_9bef2d2e7bbbb560dc05fc70","slug":"microsoft-finishes-pre-training-mai-base-2-with-2-8-trillion-parameters","model":"deepseek-v4-1-flash","headline":"Microsoft Finishes Pre-Training MAI-Base-2 With 2.8 Trillion Parameters","summary":"Microsoft says pre-training is complete for MAI-Base-2, a mixture-of-experts model with 80 billion active parameters and 2.8 trillion total. It was trained over 38 days from August to September 2026 on 16,000 GB200 GPUs, and unlike MAI-Base-1 it takes in image data, which makes up 18 percent of its 50 trillion training tokens. Post-training is underway and work on MAI-Base-3 has begun.","whyItMatters":"The model Microsoft is describing scales past MAI-Base-1 and buys its gains partly from image data folded into pre-training, so the foundation layer for its next round of products is now multimodal rather than text-only.","webCardHtml":"\u003cp\u003eThe training run itself is the number Microsoft chose to publish: 38 days across 16,000 GB200 GPUs. Microsoft says pre-training compute was cut to one twelfth while matching the pre-training score of DeepSeek-V4-Pro-Base. MAI-Base-2 also adds images at the pre-training stage, with image-related data at 18 percent of 50 trillion tokens, a change from MAI-Base-1. Microsoft says the model was not built by distillation from other models, and that the training pipeline now covers knowledge up to March 2026.\u003c/p\u003e","blueskyPost":"Pre-training is done for MAI-Base-2: 80B active parameters, 2.8T total, trained 38 days on 16,000 GB200 GPUs. Images are in the pre-training mix at 18% of 50T tokens. Post-training is running and MAI-Base-3 prep has started.","twitterPost":"Microsoft finished pre-training MAI-Base-2: 80B active parameters, 2.8T total, 38 days on 16,000 GB200 GPUs. Image data is 18% of 50T training tokens, a change from MAI-Base-1. Post-training underway; MAI-Base-3 prep has begun.","threadsPost":null,"newsletterBlurb":"Microsoft finished pre-training MAI-Base-2, an MoE model with 80 billion active parameters and 2.8 trillion total, trained over 38 days on 16,000 GB200 GPUs. Unlike MAI-Base-1, it takes in image data from the pre-training stage, at 18 percent of its 50 trillion tokens. Microsoft says it is moving ahead with post-training and preparing MAI-Base-3.","attributionJson":"[{\"source\":\"GIGAZINE\",\"url\":\"https://gigazine.net/news/20261016-microsoft-mai-base-2/\",\"title\":\"Microsoft Announces Completion of Pre-Training for 2.8 Trillion Parameter Foundation Model \\\"MAI-Base-2\\\"\"}]","lintFlagsJson":null,"lintHits":0,"costUsd":0,"inputTokens":4734,"outputTokens":660,"status":"published","repairAttempts":0,"nextRepairAt":null,"factsAttemptedAt":1792131732,"createdAt":"2026-10-16T06:17:52.000Z","publishedAt":"2026-10-16T06:21:08.000Z","updatedAt":"2026-10-16T06:21:08.000Z"},"cluster":{"id":"c_9bef2d2e7bbbb560dc05fc70","canonicalTitle":"Microsoftが2.8兆パラメータの基盤モデル「MAI-Base-2」の事前学習完了を発表","representativeArticleId":"a_31cd20d8f2ad45ada130ced8","sourceCount":1,"writtenSourceCount":1,"writeAttempts":0,"isSolo":true,"entitiesJson":"{\"anime_titles\":[],\"manga_titles\":[],\"work_titles\":[\"MAI-Base-2\"],\"studios\":[\"Microsoft\"],\"people\":[],\"type\":\"announcement\",\"domain\":\"other\",\"is_roundup\":false}","contentType":"news","status":"published","firstSeenAt":"2026-10-16T04:59:00.000Z","lastSeenAt":"2026-10-16T04:59:00.000Z","updatedAt":"2026-10-16T06:21:06.000Z"},"attribution":[{"source":"GIGAZINE","url":"https://gigazine.net/news/20261016-microsoft-mai-base-2/","title":"Microsoftが2.8兆パラメータの基盤モデル「MAI-Base-2」の事前学習完了を発表"}],"entities":{"anime_titles":[],"manga_titles":[],"work_titles":["MAI-Base-2"],"studios":["Microsoft"],"people":[],"type":"announcement","domain":"other","is_roundup":false},"keyFacts":null}