Microsoft's MAI-Thinking-1 matches top models on coding tests

Read the original article →

What happened

MAI-Thinking-1 is Microsoft's first reasoning model, shown at Build 2026 on June 2. It uses a mixture-of-experts design with 35 billion active parameters and a 256K context window. The design aims for high efficiency along with strong performance.

Microsoft says in blind tests, independent raters prefer it to Claude Sonnet 4.6. It also matches Claude Opus 4.6 on coding ability on the SWE-Bench Pro test, according to Microsoft.

Why it matters

Reasoning models think through problems in steps before answering. They tend to do better on hard math, logic, and coding tasks. Microsoft matching top labs on these tests, while keeping the model efficient, would be a real milestone for its in-house effort.

Efficiency matters for cost. A smaller active model that performs near the top can serve more users for less money.

MintedBrain take

Vendor benchmarks need a careful read, since each company picks favorable tests. Still, a competitive reasoning model from Microsoft adds pressure on rivals and more choice for builders. Run your own hard tasks through it before trusting the marketing numbers.

References

Discussion

  • Loading…

← Back to News