Nvidia GB200

Headlines · 4

1. Oct 19
   Microsoft unveils 2.8T-parameter MAI-Base-2 model

   Microsoft announced on October 15 that it completed pre-training for MAI-Base-2, a mixture-of-experts foundation model with 2.8 trillion total parameters and 80 billion active parameters. It was trained on 50 trillion tokens, including text and visual data, using 16,000 Nvidia GB200 GPUs for 38 days. Microsoft says the model outperformed the open-weight models included in its internal evaluations and estimates its training approach was 4.2 to 12.3 times more compute-efficient than DeepSeek-V4-Pro-Base when matching performance on selected tests. The company has not published equivalent comparisons against every leading proprietary model, and its efficiency figures estimate the compute required for comparable performance rather than showing universal superiority.

   TechRepublic · 🔥 0
2. Oct 19
   vLLM cuts Elastic EP scaling downtime to under 2 seconds

   vLLM published a technical post describing how asynchronous preparation and the NIXL EP backend let a running Elastic Expert Parallelism deployment add or remove GPUs with near-zero serving downtime. In its DeepSeek V4 Pro benchmark on GB200 GPUs, scaling from 16 to 20 GPUs previously paused serving for 100 seconds under the earlier Elastic EP approach and 197 seconds with a full restart, while the new flow cut the pause to 0.82 seconds; scaling back to 16 GPUs fell from 56 seconds to 1.96 seconds. The reconfiguration flow splits into background preparation — initializing joining workers, building the target communication groups, transferring weights over NVLink or RDMA, and warming up kernels and CUDA graphs — and a short commit phase that briefly pauses schedulers; vLLM says the changes are in the latest vLLM and NIXL releases.

   vLLM Blog · 🔥 0
3. Oct 15
   Prime Intellect and Nvidia post-train Nemotron 3.5 Super VL for enterprise agents

   Prime Intellect and NVIDIA built 2,500 verified training tasks aimed at teaching long-horizon enterprise agent work. According to the Prime Intellect blog, 100 GRPO steps — about 11 hours of training on 128 NVIDIA GB200 GPUs — took NVIDIA Nemotron 3.5 Super VL from 34.0% to 57.7% on AutomationBench-AA (public split), a 23.7-point gain over its starting SFT checkpoint. Each task is scored on actual changes to simulated business applications and files, with objectives the agent must complete and guardrails it must not break. The environment library, trainer and templating used for validation are open source, and the same loop runs as a hosted service on Prime Intellect Lab.

   Prime Intellect Blog · 🔥 0
4. Oct 15
   Microsoft completes MAI-Base-2 pretraining, claims it beats tested open-weight models

   Microsoft said on October 15 that it had completed pretraining of MAI-Base-2, a new base model it says outperforms all open-weight models tested. The mixture-of-experts model has 2.8 trillion total parameters and 80 billion active parameters, and was trained for 38 days on 6,000 Nvidia GB200 GPUs with 21% average floating-point utilization and 89% effective throughput, according to the company. Microsoft claims 4x to 12x the pretraining compute efficiency of DeepSeek-V4-Pro-Base and says it has begun preparing MAI-Base-3; Microsoft AI chief Mustafa Suleyman also told employees the model beats the one powering GPT-5.6 Sol using about a quarter of OpenAI's compute, though Microsoft ran those comparisons privately and OpenAI did not comment.

   IT之家 AI · 🔥 0

Experience and discussion from the community

Share my Nvidia GB200 experienceAsk about Nvidia GB200