Microsoft’s Phi-4 Reasoning Fashions Defined Merely

Microsoft isn’t like OpenAI, Google, and Meta; particularly not in terms of giant language fashions. Whereas different tech giants want to launch a number of fashions nearly overwhelming the customers with selections; Microsoft launches a number of, however these fashions at all times make it huge amongst builders around the globe. Of their newest launch, they’ve launched 2 reasoning fashions: Phi-4-Reasoning and Phi-4-Reasoning-plus, each skilled on the bottom Phi-4 mannequin. The 2 Phi-4-Reasoning fashions compete with the mighty fashions like o1, o3-mini, and DeepSeek R1. On this weblog, we’ll dive into the technical particulars, structure, coaching strategies, and efficiency of Phi-4-Reasoning fashions intimately. 

Let’s discover the Phi-4-Reasoning fashions. 

What’s Phi-4 Reasoning?

Phi-4 shouldn’t be new within the LLM world. This small and mighty language fashions broke the web when it was launched final yr. Now to cater to the rising demand for reasoning fashions, Microsoft has launched Phi-4-Reasoning fashions. These are 14 B parameters that excel at performing advanced reasoning duties involving arithmetic, coding, and STEM questions. Unline its common objective Phi-4 collection, Phi-4-Reasoning is particularly optimized for long-chain reasoning – that’s the means to interrupt down advanced multi-step issues systematically into logical steps.

Additionally Learn: Phi-4: Redefining Language Fashions with Artificial Knowledge

Phi 4 Reasoning Fashions

The 2 reasoning fashions launched by Microsoft are:

• Phi-4-Reasoning:  A reasoning mannequin skilled utilizing supervised fine-tuning or SFT on high-quality datasets. The mannequin is most well-liked for all duties that require sooner responses with guided efficiency constraints. 
• Phi-4-Reasoning-Plus: An enhanced reasoning mannequin that has been enhanced utilizing reinforcement studying or RL to enhance its efficiency however generates nearly 50% extra tokens in comparison with its counterpart. The mannequin reveals an elevated latency and therefore is really helpful for high-accuracy duties.  

The 2 14B fashions at present assist solely textual content enter, and Microsoft has launched them as open-weight so builders can freely take a look at and fine-tune them primarily based on their wants. Listed below are some key highlights of the fashions:

Key Options of Phi-4-Reasoning Fashions

For Phi-4 the group took a number of modern steps involving knowledge choice, its coaching methodology in addition to its efficiency. A few of the key issues they did had been:

Knowledge Centric Coaching 

The Knowledge Curation for coaching the Phi-4 reasoning fashions relied not simply on sheer amount however emphasised equally on the standard of information too. They particularly selected the info that was on the “edge” of the mannequin’s capabilities. This ensured that the coaching knowledge was solvable however not simply.  

The primary steps concerned in constructing the info set for Phi-4 fashions had been:

• Seed Database: The Microsft group began with publicly obtainable datasets like AIME and GPQA. These knowledge units concerned issues in algebra and geometry involving multi-step reasoning. 
• Artificial Reasoning Chains: To get complete and detailed step-by-step reasoned-out responses for the issues, the Microsoft group relied on OpenAI’s o3-mini mannequin. 

For instance, for the query “What’s the spinoff of sin(x2 )2?”; o3-mini gave the next output:

Step 1: Apply the chain rule: d/dx sin(u)=cos(u)*du/dx.  
Step 2: Let u=x² ⇒ du/dx=2x. 
Last Reply: cos(x²) * 2x.

These artificially or synthetically generated chains of well-reasoned responses gave a transparent blueprint on how a mannequin ought to construction its personal reasoning responses. 

• Choosing “Teachable Moments”: The developer group, knowingly went for prompts that challenged the bottom Phi-4 mannequin whereas being solvable. These included issues on which Phi-4 initially confirmed round 50-% accuracy. This strategy ensured that the coaching course of prevented “straightforward” knowledge that simply bolstered present patterns, and centered extra on “structured reasoning”.

The group primarily wished the Phi-4-reasoning fashions to be taught as they do, an strategy that we people normally depend on. 

Supervised Fantastic-Tuning (SFT)

Supervised Fantastic-Tuning (SFT) is the method of enhancing a pre-trained language mannequin by coaching it on rigorously chosen enter–output pairs with high-quality responses. For the Phi-4-Reasoning fashions, this meant beginning with the bottom Phi-4 mannequin after which refining it utilizing reasoning-focused duties. Basically, Phi-4-Reasoning was skilled to be taught and observe the step-by-step reasoning patterns seen in responses from o3-mini.

Coaching Particulars

• Batch Dimension: It was stored at 32. This small batch dimension allowed the mannequin to give attention to particular person examples with out being overwhelmed by the extra noise. 
• Studying charge: This was 7e-5, a reasonable charge that avoids overshooting the optimum weights throughout updates.
• Optimizer: An ordinary “Adam W” optimizer was used. This deep studying optimizer balances pace and stability. 
• Context Size:  It was taken to 32,768 tokens which was double the 16K token restrict of the bottom Phi-4 mannequin. This allowed the mannequin to deal with an extended context. 

Utilizing SFT throughout early coaching allowed the mannequin to make use of <pondering> and </pondering> tokens to separate uncooked enter from its inside reasoning. This construction made its decision-making course of clear. Additionally, the mannequin confirmed regular enhancements on the AIME benchmarks proving that the mannequin was not simply copying codecs however was constructing reasoning logic. 

Reinforcement Studying

Reinforcement studying is educating a mannequin tips on how to do higher with suggestions on all its generated outputs. The mannequin will get a reward each time it solutions accurately and is punished every time it responds incorrectly. RL was used to additional prepare the Phi-4-Reasoning -Plus mannequin. This coaching technique refined the mannequin’s math-solving expertise which evaluated the responses for accuracy and the structured strategy. 

How does RL work?

• Reward Design: The mannequin bought +1 for every appropriate response and -0.5 for incorrect response. The mannequin bought punished for repetitive phrases “Let’s see.. Let’s see..” and so on. 
• Algorithm: The GRPO algorithm or generalized reward coverage optimization algorithm, which is a variant of RL that balances exploration and exploitation was used. 
• Outcomes: Phi-4-Reasoning-Plus achieved 82.5% accuracy on AIME 2025 whereas Phi-4 Reasoning scored simply 71.4%. It confirmed improved efficiency on Omni-MATH and TSP (touring Salesman Downside) too. 

RL coaching allowed the mannequin to refine its steps iteratively and helped cut back the “hallucinations” within the generated outputs. 

Structure of Phi-4-Reasoning Fashions

The primary structure of the Phi-4-Reasoning fashions is much like the bottom Phi-4 mannequin however to assist the “reasoning” duties some key modifications had been made. 

1. The 2 placeholder tokens from Phi-4 had been repurposed. These tokens helped the mannequin to distinguish between uncooked enter and inside reasoning.
   • <pondering>: Used to mark the beginning of a reasoning block.
   • </pondering>: Used to mark the top of a reasoning block. 
2. The Phi-4 Reasoning fashions bought an prolonged context window of 32 Ok tokens to deal with the additional reasoning chains. 
3. The fashions used rotary place embeddings to raised monitor the place of tokens in lengthy sequences to assist the fashions keep coherency.
4. The fashions are skilled to work effectively on shopper {hardware} together with units like mobiles, tablets and desktops.

Phi-4-Reasoning Fashions: Benchmark Efficiency 

Phi-4-Reasoning fashions had been evaluated on numerous benchmarks to check their efficiency in opposition to totally different fashions on various duties.  

• AIME 2025: A benchmark that checks superior math, reasoning, and up to date examination problem. Phi-4-Reasoning Plus outperforms a lot of the top-performing fashions like o1, and Claude 3.7 Sonnet however continues to be behind o3-mini-high. 
• Omni-MATH: A benchmark that evaluates various math reasoning throughout matters and ranges. Each Phi-4-Reasoning and Phi-4-Reasoning plus outperform nearly all fashions solely behind DeepSeek R1.
• GPQA: A benchmark that checks mannequin efficiency on graduate-level skilled QA reasoning. The 2 Phi reasoning fashions lag behind the giants like o1, o3-mini excessive and DeepSeek R1.
• SAT: A benchmark that evaluates U.S. excessive school-level tutorial reasoning (math + verbal mix). The Phi-4-Reasoning-Plus mannequin stands among the many high 3 contenders with Phi-4-Reasoning following shut behind.
• Maze: This benchmark checks navigation + resolution pathfinding reasoning. On this benchmark, the Phi-4-reasoning fashions lag behind the top-tier fashions like o1 and Claude 3.7 sonnet.

On different benchmarks like Spatial map, TSP, and BA calendar, each the Phi-4-Reasoning fashions carry out decently. 

Additionally Learn: Find out how to Fantastic-Tune Phi-4 Regionally?

Find out how to Entry Phi-4-Reasoning Fashions?

The 2 Phi-4-Reasoning fashions can be found on Hugging Face:

Click on on the hyperlinks to move to the cuddling face web page the place you possibly can entry these fashions. On the correct facet nook of the display screen, click on on “Use This Mannequin”, click on on “Transformers” and duplicate the next code:

# Use a pipeline as a high-level helper
from transformers import pipeline

messages = [
    {"role": "user", "content": "Who are you?"},
]

pipe = pipeline("text-generation", mannequin="microsoft/Phi-4-reasoning")
pipe(messages)

Since it’s a 14B parameter mannequin and therefore requires round 40 + GB of VRAM (GPU), You’ll be able to both run these fashions on “Colab Professional” or “Runpod”. For this weblog, we ran the mannequin on “Runpod” and used “A100 GPU”. 

Set up Required Libraries

First, guarantee you’ve gotten the transformer’s library put in. You’ll be able to set up it utilizing pip:

pip set up transformers

Load the Mannequin

As soon as all of the libraries have been put in, now you can load the Phi-4-Reasoning mannequin in your pocket book:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", mannequin="microsoft/Phi-4-reasoning", max_new_tokens=4096)

Make certain to set the max_new_tokens = 4096, the mannequin generates its whole reasoning and sometimes lesser token rely can cease its output halfway.

Phi-4-Reasoning: HandsOn Purposes

We are going to now take a look at the Phi-4-reasoning fashions for 2 duties involving Logical Considering and Reasoning. Let’s begin.

Process 1: Logical Considering

Enter:

messages = [

    {"role": "user", "content": """A team is to be selected from among ten persons — A, B, C, D, E, F, G, H, I and J — subject to the following conditions.

Exactly two among E, J, l and C must be selected.

If F is selected, then J cannot be selected.

Exactly one among A and C must be selected.

Unless A is selected, E cannot be selected,

If and only if G is selected, D must not be selected.

If D is not selected, then H must be selected.

The size of a team is defined as the number of members in the team. In how many ways can the team of size 6 be selected, if it includes E? and What is the largest possible size of the team?"""

    },

]

Output:

Markdown(pipe(messages)[0]["generated_text"][1]["content"])

The mannequin thinks totally. It does an incredible job of breaking down the whole downside into small steps. The issue consists of two duties, with the given token window, it gave the reply for the primary process nevertheless it couldn’t generate the reply for the second process. What was attention-grabbing was the strategy that the mannequin took in the direction of fixing the given downside. First, it began by understanding the query, mapping out all the chances, after which it went forward into fixing every process, generally, repeating the logic that it had pre-established. 

Process 2: Clarify Working of LLMs to an 8 12 months Outdated Child

Enter:

messages = [

    {"role": "user", "content": """Explain How LLMs works by comparing their working to the photosynthesis process in a plant so that an 8 year old kid can actually understand"""

    },

]

Output:

Markdown(pipe(messages)[0]["generated_text"][1]["content"])

The mannequin hallucinates a bit whereas producing the response for this downside. Then lastly it generates the response that gives a great analogy between how LLMs work and the photosynthesis course of. It retains the language easy and at last provides a disclaimer too. 

Phi-4 Reasoning vs o3-mini: Comparability

Within the final part, we noticed how the Phi-4-Reasoning mannequin performs whereas coping with advanced issues. Now let’s evaluate its efficiency in opposition to OpenAI’s o3-mini. To do that, let’s take a look at the output generated by the 2 fashions for a similar process.

Phi-4-Reasoning

Enter:

from IPython.show import Markdown

messages = [

    {"role": "user", "content": """Suppose players A and B are playing a game with fair coins. To begin the game A and B 

    both flip their coins simultaneously. If A and B both get heads, the game ends. If A and B both get tails, they both 

    flip again simultaneously. If one player gets heads and the other gets tails, the player who got heads flips again until he 

    gets tails, at which point the players flip again simultaneously. What is the expected number of flips until the game ends?"""

    },

]

Output = pipe(messages)

Output:

Markdown(Output[0]["generated_text"][1]["content"])

o3-mini

Enter:

response = shopper.responses.create(

    mannequin="o3-mini",

    enter="""Suppose gamers A and B are taking part in a recreation with honest cash. To start the sport A and B 

    each flip their cash concurrently. If A and B each get heads, the sport ends. If A and B each get tails, they each 

    flip once more concurrently. If one participant will get heads and the opposite will get tails, the participant who bought heads flips once more till he 

    will get tails, at which level the gamers flip once more concurrently. What's the anticipated variety of flips till the sport ends?"""

)

Output:

print(response.output_text)

To verify the detailed output you possibly can confer with the next Github hyperlink.  

Outcome Analysis

Each fashions give correct solutions. Phi-4-Reasoning breaks the issue into many detailed steps and thinks by means of every one earlier than reaching the ultimate reply. o3-mini, however, combines its pondering and ultimate response extra easily, making the output clear and able to use. Its solutions are additionally extra concise and direct.

Purposes of Phi-4-Reasoning Fashions

The Phi-4-Reasoning fashions open a world of potentialities. Builders can use these fashions to develop clever techniques to cater to totally different industries. Listed below are a number of areas the place the Phi-4-Reasoning fashions can actually excel:

• Their robust efficiency in coding benchmarks (like LiveCodeBench) suggests purposes in code era, debugging, algorithm design, and automatic software program growth.
• Their means to generate detailed reasoning chains makes them well-suited for answering advanced questions that require multi-step inference and logical deduction.
• The fashions’ talents in planning duties might be leveraged in logistics, useful resource administration, game-playing, and autonomous techniques requiring sequential decision-making.
• The fashions also can contribute to designing techniques in robotics, autonomous navigation, and duties involving the interpretation and manipulation of spatial relationships.

Conclusion

The Phi-4-Reasoning fashions are open-weight and constructed to compete with high paid reasoning fashions like DeepSeek and OpenAI’s o3-mini. Since they aren’t instruction-tuned, their solutions might not at all times observe a transparent, structured format like some standard fashions, however this could enhance over time or with customized fine-tuning. Microsoft’s new fashions are highly effective reasoning instruments with robust efficiency, and so they’re solely going to get higher from right here.

Anu Madan is an professional in tutorial design, content material writing, and B2B advertising, with a expertise for reworking advanced concepts into impactful narratives. Together with her give attention to Generative AI, she crafts insightful, modern content material that educates, evokes, and drives significant engagement.