MAI-Base-2 Explained: Microsoft's 2.8T-Parameter Base Model

What Is MAI-Base-2?

MAI-Base-2 is a pre-trained foundation model from Microsoft AI (MAI), the lab led by CEO Mustafa Suleyman. A base model has finished pre-training on raw text, code, and images but has not been post-trained to follow instructions, call tools, or decline harmful requests, so it is a starting point for products rather than a product.

Suleyman announced the model on X on October 15, 2026:

The tweet rounds the size up. The technical post gives the exact figures:

• Architecture - a sparse MoE with 2.8 trillion total parameters and 80 billion active per token, with MoE in every layer and each token routed to 7 of 512 experts plus 1 shared expert.
• Native vision - pre-trained with vision from the start through an encoder-free vision-language architecture, where MAI-Base-1 was text only.
• Training data - a 50-trillion-token mixture of organic data with a March 2026 knowledge cutoff and no distillation from other models.
• Training run - 16k GB200 GPUs over 38 days in August and September 2026.
• Status - post-training is under way, and preparations for MAI-Base-3 have already started.

MAI-Base-2 is the base for Microsoft AI's next generation of models. Its predecessor followed the same path: MAI-Thinking-1, announced in June, is post-trained on top of MAI-Base-1.

How Does MAI-Base-2 Compare With MAI-Base-1?

MAI-Base-2 is larger, sparser, and trained on more data than MAI-Base-1. The specification table in the technical post shows what changed between the two:

The post reports two more changes that matter when you compare the two models:

• Recipe efficiency - progress since MAI-Base-1's release in June gave a 6.3-12.7x FLOPs-efficiency improvement, which Microsoft AI attributes to data quality and modeling choices.
• Corpus size - unique training tokens grew by roughly 36%, and vision-language data, at 4.5T unique tokens, made up 18% of the 50T-token mixture once repeated.

How Does MAI-Base-2 Score Against Open-Weight Models?

Microsoft AI compared MAI-Base-2 with frontier checkpoints available before October 2026 on four held-out evaluation sets. Scores are bits per byte (BPB), where lower is better, taken from the post's Figure 1:

Read the table with the conditions Microsoft AI attaches to it:

• Post-trained rivals - Kimi K3, Qwen3.8-2.4T-A95B, and MiMo-V2.6-Pro-RL are available only as post-trained checkpoints, and Microsoft AI cautions that post-training may affect these evaluations.
• Evaluation sources - held-out code comes from MAI's own codebase, while the STEM, math, and QA sets were commissioned from vendors to reduce the risk of training-set contamination.
• Decontamination - every evaluation set was also decontaminated against MAI's training data.
• Closest result - Kimi K3, which matches MAI-Base-2's 2.8T total parameters with more active parameters, comes within 0.0028 BPB of it on math.
• Efficiency reference - DeepSeek-V4-Pro-Base is the only released base checkpoint above one trillion parameters, so Microsoft AI used it as the reference for the 4.2-12.3x FLOPs-efficiency estimate, fitting a separate scaling law for each evaluation.

These are Microsoft AI's own held-out sets rather than public benchmarks, so outside teams cannot rerun the comparison. Treat the table as evidence about the pre-training recipe, and plan separate tests for whichever post-trained model reaches you.

What Does Bits Per Byte Measure?

Bits per byte is the average number of bits a model needs to encode each byte of held-out text, derived from the probabilities it assigns to that text. Dividing by bytes instead of tokens makes the score independent of each model's tokenizer, which is why Microsoft AI uses it to compare base models.

It is a pre-training metric, which puts it in a different category from the task benchmarks and product evals compared in LLM benchmarks, and it leaves most agent behavior unmeasured:

• What it shows - how well the base model predicts code, math, STEM, and QA text it has not seen, a direct read on how much the pre-training recipe learned.
• Instruction following - not measured, because a base model has not yet been taught to follow a system prompt or a user request.
• Tool calls - not measured, since choosing the right function with valid arguments is shaped in post-training.
• Refusals and safety - not measured. For MAI-Thinking-1, Microsoft AI trained safety in the same reinforcement learning loop as capability, after pre-training.
• Grounded answers - not measured. A lower BPB on vendor-built QA sets says nothing about whether answers stay faithful to your own documents.

A post-trained MAI model needs the checks you would run on any model change: task accuracy on your own prompts, tool-call correctness, refusal behavior, and output stability across repeated runs. The guide to LLM testing covers each of these for applications built on large language models.

How Did Microsoft AI Train MAI-Base-2?

Microsoft AI credits scaling science, model-infrastructure co-design, and scaling data for the run. Training took 38 days on 16k GB200 GPUs, and the post gives two run-level health figures:

• Model FLOPs utilization (MFU) - averaged 21%, normalized to 2.5 PFLOPS, the GB200's theoretical peak in BF16, excluding FLOPs for rematerialization and the optimizer.
• Goodput - 89%, the fraction of wall-clock time in which training made progress.

Can You Use MAI-Base-2 Today?

No. MAI-Base-2 has no public endpoint or release date. Microsoft AI says it will license its model weights to trusted partners through Foundry, but it has not named those partners.

The closest MAI model you can test now is MAI-Thinking-1, Microsoft AI's reasoning model. Its MAI-Thinking-1 post lists what the model offers in public preview on Microsoft Foundry:

• Size - a sparse MoE with 35B active and about 1T total parameters.
• Context window - 256k tokens, which Microsoft AI says is enough to fit a 600-page document.
• Developer features - function calling, layered developer instructions, and compatibility with the Chat Completions API.
• Reported results - 97.0% on AIME 2025 and 94.5% on AIME 2026, plus a SWE-Bench Pro result Microsoft AI describes as toe-to-toe with Claude Opus 4.6.

Running your evaluations on MAI-Thinking-1 now gives you a baseline on the current MAI release. When a post-trained MAI-Base-2 model arrives, you can judge it as a model swap against known scores instead of starting from zero.

How Should Teams Prepare to Test MAI Models?

Treat a post-trained MAI model as a candidate replacement for the model your agent runs on now, and gate the switch on results from your own prompts. Microsoft AI's own method is a usable template: held-out data, sets built outside the training team, and decontamination against anything the model trained on.

• Build a held-out set from your own traffic - sample real prompts and expected outcomes from logs and support tickets, and keep them out of any fine-tuning data or prompt examples.
• Baseline the current model - score your production model on that set before any swap, so the candidate has a number to beat.
• Cover what BPB skips - add tool-call cases with expected arguments, long-context cases near your real document sizes, and requests the agent must refuse or escalate.
• Repeat each prompt - run every case several times and assert invariants across runs, since one passing run proves only that run.
• Gate the switch - block the change when any metric drops below the baseline, using the gates described in LLM regression testing.

For chat and voice agents, TestMu AI's Agent Testing runs this comparison as a regression suite. It generates 60-100+ test scenarios from your PRDs and documents, scores each response on 9 quality metrics, including hallucination, bias, completeness, and context awareness, and shows the metric deltas between a run on your current model and a run on the candidate.

Each run ends in a production-readiness verdict: Green at 80 or above, Yellow from 65 to 79, and Red below 65, as the agent features and metrics docs define it. A model update that improves one metric and degrades another shows up in the per-metric deltas, even when the overall verdict holds.

If you plan to use MAI-Thinking-1 or a later MAI model as a coding agent, check what it builds as well as what it reports. Kane CLI runs a plain-English objective against the rendered app in a real browser and returns an evidence-backed pass or fail that a CI step, or the coding agent itself, can read.

Note: Baseline your agent on the model it runs today, then rerun the same suite on every candidate model with TestMu AI Agent Testing. Create a free TestMu AI account!

Getting Started With MAI Model Evaluation

Pick the agent most likely to move to an MAI model, export a set of real prompts with their expected outcomes, and run them through Agent Testing on the model it uses today. That baseline is the score a post-trained MAI-Base-2 release will have to beat, and the guide to testing your first AI agent walks through the first run.