Viewing enriched Twitter post
PyTorch and Nebius have collaborated to speed-up pre-training DeepSeek-V3 (16B & 671B) on 256 NVIDIA B200 GPUs using TorchTitan. Combining MXFP8 via TorchAO with DeepEP yielded up to 41% faster training throughput over BF16 (with equivalent convergence). Full results and reproducible configs in the blog: š https://t.co/udvt6W86FM ā Alireza Shamsoshoara, Matthias Reso, Hamid Shojanazeri, @vega_myhre, Hooman Ramezani #PyTorch #OpenSourceAI #TorachAO #MXFP8
{
"media": [
{
"type": "photo",
"url": "https://crmoxkoizveukayfjuyo.supabase.co/storage/v1/object/public/media/posts/2036884532104736961/media_0.jpg",
"filename": "media_0.jpg"
}
],
"processed_at": "2026-03-25T19:21:21.997681",
"pipeline_version": "2.0"
}{
"type": "tweet",
"id": "2036884532104736961",
"url": "https://x.com/PyTorch/status/2036884532104736961",
"twitterUrl": "https://twitter.com/PyTorch/status/2036884532104736961",
"text": "PyTorch and Nebius have collaborated to speed-up pre-training DeepSeek-V3 (16B & 671B) on 256 NVIDIA B200 GPUs using TorchTitan. \n\nCombining MXFP8 via TorchAO with DeepEP yielded up to 41% faster training throughput over BF16 (with equivalent convergence). Full results and reproducible configs in the blog:\n\nš https://t.co/udvt6W86FM\n\nā Alireza Shamsoshoara, Matthias Reso, Hamid Shojanazeri, @vega_myhre, Hooman Ramezani\n\n#PyTorch #OpenSourceAI #TorachAO #MXFP8",
"source": "Twitter for iPhone",
"retweetCount": 0,
"replyCount": 1,
"likeCount": 5,
"quoteCount": 0,
"viewCount": 324,
"createdAt": "Wed Mar 25 19:14:50 +0000 2026",
"lang": "en",
"bookmarkCount": 1,
"isReply": false,
"inReplyToId": null,
"conversationId": "2036884532104736961",
"displayTextRange": [
0,
277
],
"inReplyToUserId": null,
"inReplyToUsername": null,
"author": {
"type": "user",
"userName": "PyTorch",
"url": "https://x.com/PyTorch",
"twitterUrl": "https://twitter.com/PyTorch",
"id": "776585502606721024",
"name": "PyTorch",
"isVerified": false,
"isBlueVerified": true,
"verifiedType": null,
"profilePicture": "https://pbs.twimg.com/profile_images/1813965160702451712/yXV1vRhr_normal.jpg",
"coverPicture": "https://pbs.twimg.com/profile_banners/776585502606721024/1761575044",
"description": "",
"location": "",
"followers": 481994,
"following": 83,
"status": "",
"canDm": false,
"canMediaTag": true,
"createdAt": "Fri Sep 16 00:56:26 +0000 2016",
"entities": {
"description": {
"urls": []
},
"url": {}
},
"fastFollowersCount": 0,
"favouritesCount": 860,
"hasCustomTimelines": true,
"isTranslator": false,
"mediaCount": 1362,
"statusesCount": 3116,
"withheldInCountries": [],
"affiliatesHighlightedLabel": {},
"possiblySensitive": false,
"pinnedTweetIds": [],
"profile_bio": {
"description": "Tensors and neural networks in Python with strong hardware acceleration. PyTorch is an open source project at the Linux Foundation. #PyTorchFoundation",
"entities": {
"description": {
"hashtags": [
{
"indices": [
132,
150
],
"text": "PyTorchFoundation"
}
],
"symbols": [],
"urls": [],
"user_mentions": []
},
"url": {
"urls": [
{
"display_url": "pytorch.org",
"expanded_url": "http://pytorch.org",
"indices": [
0,
23
],
"url": "https://t.co/6SwTBhUwTJ"
}
]
}
}
},
"isAutomated": false,
"automatedBy": null
},
"extendedEntities": {
"media": [
{
"display_url": "pic.twitter.com/bi2Z0nnZ4s",
"expanded_url": "https://twitter.com/PyTorch/status/2036884532104736961/photo/1",
"ext_media_availability": {
"status": "Available"
},
"features": {
"large": {
"faces": [
{
"h": 63,
"w": 63,
"x": 172,
"y": 363
}
]
},
"orig": {
"faces": [
{
"h": 63,
"w": 63,
"x": 172,
"y": 363
}
]
}
},
"id_str": "2036884511368126464",
"indices": [
278,
301
],
"media_key": "3_2036884511368126464",
"media_results": {
"id": "QXBpTWVkaWFSZXN1bHRzOgwAAQoAARxEd6vUl6AACgACHER3sKiXMMEAAA==",
"result": {
"__typename": "ApiMedia",
"id": "QXBpTWVkaWE6DAABCgABHER3q9SXoAAKAAIcRHewqJcwwQAA",
"media_key": "3_2036884511368126464"
}
},
"media_url_https": "https://pbs.twimg.com/media/HER3q9SXoAA_Qxv.jpg",
"original_info": {
"focus_rects": [
{
"h": 1075,
"w": 1920,
"x": 0,
"y": 0
},
{
"h": 1080,
"w": 1080,
"x": 0,
"y": 0
},
{
"h": 1080,
"w": 947,
"x": 55,
"y": 0
},
{
"h": 1080,
"w": 540,
"x": 258,
"y": 0
},
{
"h": 1080,
"w": 1920,
"x": 0,
"y": 0
}
],
"height": 1080,
"width": 1920
},
"sizes": {
"large": {
"h": 1080,
"w": 1920
}
},
"type": "photo",
"url": "https://t.co/bi2Z0nnZ4s"
}
]
},
"card": null,
"place": {},
"entities": {
"hashtags": [
{
"indices": [
424,
432
],
"text": "PyTorch"
},
{
"indices": [
433,
446
],
"text": "OpenSourceAI"
},
{
"indices": [
447,
456
],
"text": "TorachAO"
},
{
"indices": [
457,
463
],
"text": "MXFP8"
}
],
"symbols": [],
"urls": [
{
"display_url": "pytorch.org/blog/enabling-ā¦",
"expanded_url": "https://pytorch.org/blog/enabling-up-to-41-faster-pre-training-mxfp8-and-deepep-for-deepseek-v3-on-b200-with-torchtitan/",
"indices": [
311,
334
],
"url": "https://t.co/udvt6W86FM"
}
],
"user_mentions": [
{
"id_str": "1906137662827937792",
"indices": [
394,
405
],
"name": "Daniel Vega-Myhre",
"screen_name": "vega_myhre"
}
]
},
"quoted_tweet": null,
"retweeted_tweet": null,
"isLimitedReply": false,
"article": null
}