Free tools Windows power users keep installed
One-click scans. No signup required.
iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Microsoft AI says it has completed pre-training MAI-Base-2, a mixture-of-experts (MoE) foundation model with 2.8 trillion total parameters and 80 billion active parameters per token. In Microsoft’s held-out evaluations, the company says the model beat every open-weight model it tested and required an estimated 4.2–12.3 times fewer pre-training FLOPs than DeepSeek-V4-Pro-Base to reach comparable results. MAI-Base-2 is still in post-training, so it is not yet publicly downloadable or generally available.
What Microsoft announced
Microsoft AI published its announcement on October 15, 2026. The company describes MAI-Base-2 as a new state-of-the-art base model rather than a finished consumer chatbot. Its headline comparison is against open-weight models on held-out evaluations, using bits-per-byte (BPB), a tokenizer-agnostic measure of how efficiently a model predicts data.
Microsoft says MAI-Base-2 “outperforms every open-weight model we tested on our held-out evaluations.” That is a company-reported result, not an independently reproduced benchmark. The announcement does not include public weights, inference prices, or a general release date.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →How big is MAI-Base-2?
MAI-Base-2 has 2.8 trillion total parameters, but only about 80 billion are activated for each token. This is the defining MoE trade-off: the model stores a very large collection of expert weights while routing each token through a smaller subset. Total parameters indicate the model’s overall capacity and memory footprint; active parameters are more relevant to the computation required for an individual token.
#1 Best Overall
- ATTENTION BUYERS: EXPORTING this item outside the U.S. requires strict compliance with the Bureau of Industry and Security (BIS) U.S. Export Administration Regulations (EAR). Non-compliance carries fines up to $1 MILLION PER VIOLATION AND UP TO 20 YEARS IMPRISONMENT for both the buyer and seller. To uphold our regulatory obligations, we strictly follow BIS "Know Your Customer" guidance by screening each order and immediately aborting transactions showing unresolved regulatory red flags. ***
- [Double-Flow-Through Design] The RTX PRO 6000 Blackwell features a double-flow-through cooling design, optimizing efficiency and airflow to sustain peak performance under 600W power loads. | [5th Gen Tensor Cores] Deliver up to 3X the performance of the previous generation and support for FP4 precision for faster AI model processing times with reduced memory usage, enabling local fine-tuning of LLMs and generative AI | [4th Gen Ray Tracing Cores] Double the ray-triangle intersection rate of the previous generation to create photoreal, physically accurate scenes and immersive 3D designs with RTX Mega Geometry, which enables up to 100X more ray-traced triangles.
- [PCIe Gen 5] Support for PCIe Gen 5 provides double the bandwidth of PCIe Gen 4, improving data-transfer speeds from CPU memory and unlocking faster performance for data-intensive tasks like AI, data science, and 3D modeling. | [GDDR7 Memory] With 96 GB of GPU memory and 1.8 TB ps bandwidth, it can tackle massive 3D and AI projects, fine-tune AI models locally, explore large-scale VR environments, and drive larger multi-app workflows.
- [DisplayPort 2.1] Achieve unparalleled visual clarity and performance, driving high resolution displays at up to 8K at 240 Hz and 16K at 60 Hz. Increased bandwidth enables seamless multi-monitor setups while HDR and higher color depth support ensures superior color accuracy for precision work, such as video editing, 3D design, and live broadcasting.
- [NVIDIA Blackwell Streaming Multiprocessor] The new SM features increased processing throughput, and new neural shaders that integrate neural networks inside of programmable shaders | DLSS 4: Multi Frame Generation ensures ultra-smooth frame pacing for lifelike simulations. | [Universal MIG] Divide a single RTX PRO 6000 Blackwell into multiple isolated instances, each with dedicated resources, allowing for concurrent execution of multiple workloads, optimized GPU utilization, and secure isolation of different applications or users. [WARRANTY] 3 YR Manufacturer's Warranty. Bulk OEM Packaging. Retail Packaging is NOT included.
| Specification | MAI-Base-2 |
|---|---|
| Architecture | Mixture of experts (MoE) foundation model |
| Total parameters | 2.8 trillion |
| Active parameters per token | 80 billion |
| Training mixture | 50 trillion tokens |
| Vision-language portion | 4.5 trillion unique vision-language tokens; 18% of the mixture |
| Knowledge cutoff | March 2026 |
The 80-billion active count should not be read as a claim that the model can be run with only 80 billion parameters resident in memory. Serving an MoE model still requires storing or distributing its full expert set, although routing reduces the arithmetic performed for each token.
Why Microsoft attributes the result to its training approach
Microsoft groups the work into three initiatives: scaling ladders with source scaling, co-design of the model and its infrastructure, and larger data pipelines that include vision data.
More efficient MoE execution
The company reports that MoE-layer model FLOPs utilization (MFU) improved from 24% to 35%. That gain allowed Microsoft to use MoE layers throughout the network instead of limiting them to selected parts. Microsoft also describes multi-round dropless MoE routing, which avoids dropping tokens when expert capacity is reached.
Recommended Free Tools
Rank #2
- Professional GPU with Blackwell Architecture
- Blackwell Architecture
- 24GB GDDR7 with PCIe 5.0 & Ray Tracing
- AI Workstation
Optimization and stability techniques
The training recipe used Muon optimization and activation-gradient clipping. Microsoft presents these as parts of a broader effort to keep very large, sparse training runs stable and computationally productive; it does not isolate the contribution of each technique.
Data and modality
The 50-trillion-token mixture included 4.5 trillion unique vision-language tokens, with vision-language data representing 18% of the total. Microsoft says its pre-training used “organic data which does not involve distillation from other models.” The stated knowledge cutoff is March 2026, so information created after that month would not be expected to be part of the pre-training corpus.
What the training run required
Microsoft says the run used 16,000 GB200 GPUs for 38 days during August and September 2026. Average MFU was 21%, and reported goodput was 89%.
Rank #3
- Professional GPU with Blackwell Architecture in Compact Small Form Factor (SFF)
- Blackwell Architecture
- 24GB GDDR7 with PCIe 5.0 & Ray Tracing
- AI Workstation
- Hardware: 16,000 GB200 GPUs.
- Elapsed training time: 38 days in August and September 2026.
- Average MFU: 21% across the run.
- Goodput: 89%, reflecting the portion of time producing useful training work after operational overhead.
These figures describe Microsoft’s specific infrastructure and run, not a minimum hardware requirement for every future deployment. The announcement does not state the model’s inference memory requirement or a recommended server configuration.
How MAI-Base-2 compares with the models Microsoft tested
Microsoft compared the base model with DeepSeek-V4-Pro-Base, Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL. The company reports lower BPB for MAI-Base-2 on the held-out tests, but it does not publish a single aggregate score in the announcement.
| Comparison axis | MAI-Base-2 | What Microsoft discloses about the comparison |
|---|---|---|
| Held-out text metric | Lower BPB than the open-weight models tested, according to Microsoft | BPB is tokenizer-agnostic; detailed per-test values are not stated here |
| Pre-training efficiency | Estimated 4.2–12.3× advantage versus DeepSeek-V4-Pro-Base | Estimate fitted separately for four evaluations |
| Evaluation contamination | Sets were decontaminated against MAI-Base-2 training data | Microsoft reports the same decontamination approach for the comparison |
| Code evaluation | Microsoft’s own codebase | Not an independent public code benchmark |
| STEM, math and Q&A | Vendor-commissioned evaluations | Vendor names and full test details are not stated in the announcement |
| Post-training status of comparison models | Base-model result | Microsoft warns that post-training may affect results for Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL |
What “4.2–12.3× more FLOPs-efficient” means
The range is an estimate of pre-training compute efficiency, not a claim that every prompt runs 4.2 to 12.3 times faster or costs that much less to serve. Microsoft fitted an independent scaling law relating BPB to training FLOPs for each of four held-out evaluations. It then estimated how much compute its own recipe would need to match DeepSeek-V4-Pro-Base on those evaluations.
Rank #4
- NVIDIA Ampere Architecture-based CUDA Cores - Double-speed processing for single-precision floating point (FP32) operations and improved power efficiency provide significant performance improvements for graphics and simulation workflows, such as complex 3D computer-aided design (CAD) and computer-aided engineering (CAE), on the desktop.
- Second-Generation RT Cores - With up to 2X the throughput over the previous generation and the ability to concurrently run ray tracing with either shading or denoising capabilities, second-generation RT Cores deliver massive speedups for workloads like photorealistic rendering of movie content, architectural design evaluations, and virtual prototyping of product designs. This technology also speeds up the rendering of ray-traced motion blur for faster results with greater visual accuracy.
- Third-Generation Tensor Cores - New Tensor Float 32 (TF32) precision provides up to 5X the training throughput over the previous generation to accelerate AI and data science model training without requiring any code changes. Hardware support for structural sparsity doubles the throughput for inferencing. Tensor Cores also bring AI to graphics with capabilities like DLSS, AI denoising, and enhanced editing for select applications.
- Third-Generation NVIDIA NVLink - Increased GPU-to-GPU interconnect bandwidth provides a single scalable memory to accelerate graphics and compute workloads and tackle larger datasets.
- 48 Gigabytes (GB) of GPU Memory - Ultra-fast GDDR6 memory, scalable up to 96 GB with NVLink, gives data scientists, engineers, and creative professionals the large memory necessary to work with massive datasets and workloads like data science and simulation.
The resulting estimates span 4.2× to 12.3×, which is why the announcement also summarizes the result as “4–12×.” The calculation depends on the fitted scaling laws, the selected evaluations and the DeepSeek comparison point. It is therefore a Microsoft analysis rather than an independently verified efficiency benchmark.
Can you download or run MAI-Base-2?
Not at publication. Microsoft says MAI-Base-2 was undergoing post-training on October 15, 2026. It plans to license the weights to trusted partners through Microsoft Foundry, but it has not announced a public download, open-weight release, general availability date, inference pricing or public signup process.
For now, the practical access path is to wait for Microsoft’s Foundry partner program and its eventual eligibility terms. “Trusted partners” indicates a controlled licensing route, not an unrestricted model release.
What remains unknown
- Whether post-training will improve or change the reported held-out results.
- The model’s final chat, instruction-following and safety behavior.
- Public weights, license text and redistribution rights.
- Context length, supported input and output formats, and production latency.
- Foundry regions, pricing, quotas and a general availability date.
- Independent reproductions of the BPB results or the 4.2–12.3× efficiency estimate.
Why the announcement matters
MAI-Base-2 combines unusually high total capacity with a much smaller active-per-token budget, while Microsoft claims a substantial pre-training efficiency lead over the open base model it used as its principal reference. If later evaluations and independent testing support those claims, the approach could make trillion-parameter capacity more practical to train. Until weights, post-training results and external measurements appear, the strongest defensible conclusion is narrower: Microsoft has reported an impressive internal comparison for a model that is not yet generally available.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
