Microsoft AI said it will make its in-house "MAI" models, including reasoning models, fine-tunable on Fireworks AI's platform, aiming to let developers and enterprises fully own their data and custom models.
June 3 · Microsoft AI × Fireworks AI
Microsoft Opens Its MAI Models to Developer Fine-Tuning
A new partnership with Fireworks AI lets enterprises customize Microsoft's in-house MAI models on their own data — and keep the resulting weights "fully yours." It's the first place developers can tune the weights themselves.
7
In-house MAI models unveiled at Build 2026
5B
Parameters in the lightweight MAI-Code-1-Flash
10×
Lower cost / higher efficiency from "Frontier Tuning"
"Frontier Tuning": small, specialized, cheaper
A tuned MAI model hit GPT-5.4-class performance — at a fraction of the cost.
10× cost
general large model
Same task, same class of result — up to 10× more efficient with a smaller, workflow-tuned model.
The MAI lineup
MAI-Thinking-1
Reasoning flagship · mid-size
Top-tier on software-engineering benchmarks; on par with or better than Sonnet 4.6 in human evals. Trained on clean data, no distillation.
MAI-Code-1-Flash
5B parameters
Inexpensive, Haiku-class. Aimed at GitHub Copilot and VS Code integration.
MAI-Image-2.5
Image generation
Noted for strong safety filtering across generated content.
Why it matters
First environment to tune MAI weights yourself
Resulting weights stay "fully yours"
Reduces Microsoft's reliance on OpenAI
Fireworks adds fast inference + BYOW in Azure Foundry
Still to prove
MAI-specific fine-tuning is "coming soon"
Real-world tuning reports remain scarce
MAI-Image-2.5 filter can over-reject harmless prompts
Some models score below DeepSeek V4 on benchmarks
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…