At present, we’re excited to announce two main enhancements to mannequin fine-tuning in Azure AI Foundry—Reinforcement Nice-Tuning (RFT) with o4-mini, coming quickly, and Supervised Nice-Tuning (SFT) for the 4.1-nano mannequin, out there now.
At present, we’re excited to announce three main enhancements to mannequin fine-tuning in Azure AI Foundry—Reinforcement Nice-Tuning (RFT) with o4-mini (coming quickly), Supervised Nice-Tuning (SFT) for the GPT-4.1-nano and Llama 4 Scout mannequin (out there now). These updates replicate our continued dedication to empowering organizations with instruments to construct extremely custom-made, domain-adapted AI programs for real-world affect.
With these new fashions, we’re unblocking two main avenues of LLM customization: GPT-4.1-nano is a robust small mannequin, preferrred for distillation, whereas o4-mini is the primary reasoning mannequin you’ll be able to fine-tune, and Llama 4 Scout is a best-in-class open supply mannequin.
Reinforcement Nice-Tuning with o4-mini
Reinforcement Nice-Tuning introduces a brand new stage of management for aligning mannequin habits with complicated enterprise logic. By rewarding correct reasoning and penalizing undesirable outputs, RFT improves mannequin decision-making in dynamic or high-stakes environments.
Coming quickly for the o4-mini mannequin, RFT unlocks new prospects to be used instances requiring adaptive reasoning, contextual consciousness, and domain-specific logic—all whereas sustaining quick inference efficiency.
Actual world affect: DraftWise
DraftWise, a authorized tech startup, used reinforcement fine-tuning (RFT) in Azure AI Foundry Fashions to boost the efficiency of reasoning fashions tailor-made for contract era and evaluation. Confronted with the problem of delivering extremely contextual, legally sound ideas to attorneys, DraftWise fine-tuned Azure OpenAI fashions utilizing proprietary authorized information to enhance response accuracy and adapt to nuanced person prompts. This led to a 30% enchancment in search outcome high quality, enabling attorneys to draft contracts sooner and deal with high-value advisory work.
Reinforcement fine-tuning on reasoning fashions is a possible sport changer for us. It’s serving to our fashions perceive the nuance of authorized language and reply extra intelligently to complicated drafting directions, which guarantees to make our product considerably extra helpful to attorneys in actual time.
—James Ding, founder and CEO of DraftWise.
When must you use Reinforcement Nice-Tuning?
Reinforcement Nice-Tuning is greatest suited to use instances the place adaptability, iterative studying, and domain-specific habits are important. You need to take into account RFT in case your state of affairs entails:
- Customized Rule Implementation: RFT thrives in environments the place determination logic is very particular to your group and can’t be simply captured by static prompts or conventional coaching information. It permits fashions to be taught versatile, evolving guidelines that replicate real-world complexity.
- Area-Particular Operational Requirements: Superb for situations the place inside procedures diverge from trade norms—and the place success relies on adhering to these bespoke requirements. RFT can successfully encode procedural variations, equivalent to prolonged timelines or modified compliance thresholds, into the mannequin’s habits.
- Excessive Determination-Making Complexity: RFT excels in domains with layered logic and variable-rich determination timber. When outcomes depend upon navigating quite a few subcases or dynamically weighing a number of inputs, RFT helps fashions generalize throughout complexity and ship extra constant, correct choices.
Instance: Wealth advisory at Contoso Wellness
To showcase the potential of RFT, take into account Contoso Wellness, a fictitious wealth advisory agency. Utilizing RFT, the o4-mini mannequin realized to adapt to distinctive enterprise guidelines, equivalent to figuring out optimum consumer interactions primarily based on nuanced patterns just like the ratio of a consumer’s web value to out there funds. This enabled Contoso to streamline their onboarding processes and make extra knowledgeable choices sooner.
Supervised Nice-Tuning now out there for GPT-4.1-nano
We’re additionally bringing Supervised Nice-Tuning (SFT) to the GPT-4.1-nano mannequin—a small however highly effective basis mannequin optimized for high-throughput, cost-sensitive workloads. With SFT, you’ll be able to instill your mannequin with company-specific tone, terminology, workflows, and structured outputs—all tailor-made to your area. This mannequin will probably be out there for fine-tuning within the coming days.
Why Nice-tune GPT-4.1-nano?
- Precision at Scale: Tailor the mannequin’s responses whereas sustaining velocity and effectivity.
- Enterprise-Grade Output: Guarantee alignment with enterprise processes and tone-of-voice.
- Light-weight and Deployable: Excellent for situations the place latency and price matter—equivalent to customer support bots, on-device processing, or high-volume doc parsing.
In comparison with bigger fashions, 4.1-nano delivers sooner inference and decrease compute prices, making it nicely suited to large-scale workloads like:
- Buyer help automation, the place fashions should deal with hundreds of tickets per hour with constant tone and accuracy.
- Inside data assistants that observe firm type and protocol in summarizing documentation or responding to FAQs.
As a small, quick, however extremely succesful mannequin, GPT-4.1-nano makes an excellent candidate for distillation as nicely. You should utilize fashions like GPT-4.1 or o4 to generate coaching information—or seize manufacturing site visitors with saved completions—and train 4.1-nano to be simply as sensible!
Llama 4 Nice-Tuning now out there
We’re additionally excited to announce help for fine-tuning Meta’s Llama 4 Scout—a leading edge,17 billion energetic parameter mannequin which gives an trade main context window of 10M tokens whereas becoming on a single H100 GPU for inferencing. It’s a best-in-class mannequin, and extra highly effective than all earlier era llama fashions.
Llama 4 fine-tuning is obtainable in our managed compute providing, permitting you to fine-tune and inference utilizing your personal GPU quota. Obtainable in each Azure AI Foundry and as Azure Machine Studying elements, you might have entry to extra hyperparameters for deeper customization in comparison with our serverless expertise.
Get began with Azure AI Foundry at present
Azure AI Foundry is your basis for enterprise-grade AI tuning. These fine-tuning enhancements unlock new frontiers in mannequin customization, serving to you construct clever programs that suppose and reply in ways in which replicate your online business DNA.
- Use Reinforcement Nice-tuning with o4-mini to construct reasoning engines that be taught from expertise and evolve over time. Coming quickly in Azure AI Foundry, with regional availability for East US2 and Sweden Central.
- Use Supervised Nice-Tuning with 4.1-nano to scale dependable, cost-efficient, and extremely custom-made mannequin behaviors throughout your group. Obtainable now in Azure AI Foundry in North Central US and Sweden Central.
- Attempt Llama 4 scout high-quality tuning to customise a best-in-class open supply mannequin. Obtainable now in Azure AI Foundry mannequin catalog and Azure Machine Studying.
With Azure AI Foundry, fine-tuning isn’t nearly accuracy—it’s about belief, effectivity, and flexibility at each layer of your stack.
Discover additional:
We’re simply getting began. Keep tuned for extra mannequin help, superior tuning methods, and instruments that can assist you construct AI that’s smarter, safer, and uniquely yours.
