Microsoft Phi-4 in 2026: The Open-Weight Small Model That Runs on Your Laptop

TL;DR: Microsoft's Phi-4 family is the small-language-model line that keeps embarrassing models several times its size — and it's MIT-licensed and free to download. The lineup spans Phi-4-mini (3.8B), Phi-4-multimodal (5.6B, text+vision+speech), the flagship Phi-4 (14B), the reasoning pair Phi-4-reasoning / Phi-4-reasoning-plus (14B), and the newest member, Phi-4-reasoning-vision-15B, shipped March 4, 2026, which decides on its own when to "think" and when to answer instantly. Base Phi-4 posts MMLU 84.8, GPQA 56.1, MATH 80.4, and HumanEval 82.6 — outscoring its own teacher (GPT-4o) on GPQA and MATH. The practical headline: a Q4 quant of the 14B runs in roughly 8-10GB of VRAM, so this is frontier-adjacent reasoning on a gaming laptop for $0 in API spend. There is no GA "Phi-5" as of June 2026. Here's the honest state of the family and what it means for builders.

What Is the Phi Family in June 2026?

Phi is Microsoft Research's bet that data quality beats parameter count. The whole program is built on "textbook-quality" curated and synthetic training data rather than scraping more of the internet, and the result is a family of small models that punch well above their weight class.

As of June 2026, the current generation is the Phi-4 family — there is no generally available Phi-5. (You'll find "Phi-5" deployment guides floating around; treat those as speculative or pre-release until Microsoft ships an official model card. I'm not going to feature a model that doesn't have one.) The family is open-weight under the MIT license, which is about as permissive as it gets: use it commercially, fine-tune it, redistribute it, no royalty.

Five things define where the family sits right now:

1. It's a small family on purpose. The largest member is 15B parameters. Nothing here competes with a frontier model on raw breadth — the entire pitch is capability-per-parameter and the ability to run on hardware you already own.

2. The newest model is multimodal and adaptive. Phi-4-reasoning-vision-15B (March 2026) is trained to default to fast, direct inference on perception tasks and only spend tokens on long chain-of-thought when the problem — math, science, diagrams — actually needs it.

3. Reasoning came to the small tier. Phi-4-reasoning and Phi-4-reasoning-plus (April 2025, 14B, MIT) brought o1-mini-class math performance to a model you can run locally.

4. The economics are inverted versus the API world. Open weights mean the per-token price is $0. Your cost is hardware and electricity. For high-volume, privacy-sensitive, or offline workloads, that changes the math entirely.

5. It runs everywhere small models run. Ollama, LM Studio, llama.cpp, ONNX Runtime, Azure AI Foundry, Foundry Local, and Hugging Face all carry official builds.

What Are the Current Variants?

Here's the full Phi-4 lineup with the specifics that matter for self-hosting. VRAM figures are for the commonly used Q4_K_M GGUF quant on the 14B/15B models and are corroborated across community runner guides; treat them as practical estimates, not guarantees.

A few notes the table can't hold. Phi-4-mini carries a 200K-token vocabulary for stronger multilingual coverage, grouped-query attention for efficient long-context generation, and built-in function calling — it's the one to reach for when you want a tiny, fast, tool-using model. Phi-4-multimodal was the family's first model to fuse text, vision, and speech in one set of weights, and specializes in speech recognition, translation, and audio Q&A. The base Phi-4's 16K context is its most notable limitation versus the 128K minis — if you need long-context, the mini or multimodal variants are the better fit despite being smaller.

What Are the Verified Benchmarks?

A sourcing note first, because it matters. The base Phi-4 numbers come from Microsoft's Phi-4 Technical Report run through OpenAI's simple-evals framework and are cross-referenced against llm-stats and the Hugging Face model card. Where a figure is from Microsoft's own evals and not yet broadly third-party reproduced, I mark it vendor-claimed. Reasoning and vision scores in particular lean on Microsoft's reporting — treat them accordingly.

Base Phi-4 (14B) — the small model punching up

The standout context: on GPQA and MATH, Phi-4 outscores GPT-4o — the very model that generated much of its synthetic training data. A 14B model beating its multi-hundred-billion-parameter teacher on graduate science and competition math is the entire Phi thesis in two numbers. On HumanEval it posted the strongest coding score of any open-weight model in its size class at release.

Phi-4-reasoning / reasoning-plus (14B) — local o1-mini-class math

These are vendor-claimed from Microsoft's reasoning technical report, but the comparison is striking: a 14B model trading blows with — and on AIME, beating — 70B-parameter distillations and o1-mini. The reasoning variant was fine-tuned on ~8.3B tokens of synthetic chain-of-thought traces generated by o3-mini; the "plus" variant added a short GRPO reinforcement-learning phase. Both are MIT-licensed and run locally on a single consumer GPU.

Phi-4-reasoning-vision-15B — adaptive multimodal (March 2026)

All vendor-claimed from the Phi-4-reasoning-vision tech report. The honest read: this model is strong on math, science diagrams, and UI/chart grounding and noticeably weaker on general multimodal understanding (MMMU 54.3 trails larger vision models like Qwen 3.5-class systems, which clear 80+). It was trained on just 200B multimodal tokens — versus the 1T+ that Qwen 2.5/3 VL, Kimi-VL, and Gemma 3 used — and Microsoft's framing is that it matches "much slower models that require ten times or more compute-time" on its strong domains. Believe the targeted strengths; don't expect it to be a general-purpose vision model.

How Does It Compare to Other Small / Open Models?

Where the Phi-4 base model sits against the small/open field. Numbers are best-effort from each model's reporting; cross-model benchmark comparisons always carry methodology caveats, so read the shape, not the third decimal.

Two honest caveats. First, Phi-4's base 16K context is short next to Gemma 3 and Qwen's 128K — if long-context is your bottleneck, Phi-4-mini (128K) or a different family wins. Second, on general multimodal breadth, Gemma 3 and the Qwen VL line are stronger than Phi-4-reasoning-vision; Phi's vision model is a specialist, not a generalist. Where Phi wins decisively is STEM reasoning density per parameter and per gigabyte of VRAM — and the MIT license, which is more permissive than Gemma's or Llama's. For the broader open-local landscape, see the best open local LLMs guide and the Gemma 3 deep-dive.

What's the Self-Host Story? (VRAM, Quantization, Runners)

This is the part that actually matters for a small open model, and it's where Phi-4 shines. You are not paying per token. You are paying for a GPU you probably already have.

VRAM by quant (Phi-4 14B):

Community testing puts Q4_K_M at ~95% of full-precision quality on reasoning tasks, which is why it's the default recommendation. The 3.8B Phi-4-mini drops to roughly 3-4GB at Q4 — it runs comfortably on an 8GB laptop GPU, an integrated GPU with enough shared memory, or even CPU-only at reduced speed.

Where to run it:

• Ollama — ollama run phi4. Official GGUFs ship in current releases; the same file works in llama.cpp, KoboldCPP, and friends.
• LM Studio — official phi4 GGUFs, GUI, one-click download. Best for non-terminal users.
• ONNX Runtime / Foundry Local — Microsoft's own path for optimized on-device inference, including Windows CPU/GPU/NPU.
• Azure AI Foundry — managed endpoints if you'd rather not self-host but still want the model.
• NVIDIA NIM — Phi-4-mini is packaged as a NIM microservice for production deployment.
• Hugging Face — raw weights for fine-tuning and custom pipelines.

The same GGUF artifact is portable across the whole local ecosystem, so you're not locked into one runner. For a privacy-first local setup walkthrough, the Ollama local AI guide covers the mechanics.

What Does On-Device / Edge Economics Actually Look Like?

The open-weight angle isn't a footnote — it's the whole value proposition for a small model. Here's the calculus.

Cost. Per-token price is $0. A workload that would cost real money against a hosted API — say, classifying or summarizing millions of documents — costs you electricity and the amortized price of a GPU. Once the hardware is bought, the marginal cost of inference is effectively zero. For high-volume, repetitive tasks, that flips the build-vs-buy decision hard toward self-hosting.

Privacy. Nothing leaves the machine. For regulated data — health, legal, financial, internal source code — a model that runs entirely on-prem or on-device sidesteps an entire category of data-governance problems. No API logs, no third-party data processing agreements, no exfiltration surface.

Latency and offline. No network round-trip. A 3.8B Phi-4-mini on a laptop NPU responds locally, works on a plane, and degrades gracefully without connectivity. For embedded and edge deployments — kiosks, IoT, field devices — this is the difference between feasible and not.

The honest tradeoff. You give up frontier breadth. Phi-4 will not match Claude Opus or Gemini on the hardest, broadest tasks, and it never claimed to. The discipline is matching the model to the task's difficulty and cost-of-error: route the genuinely hard, high-stakes work to a frontier API, and run the high-volume, well-scoped, privacy-sensitive work on Phi-4 locally. That's the same routing logic the FrankX models tracker applies across the whole field.

What Does It Mean for Builders?

For developers and edge deployments

If you have a task that's well-scoped — structured extraction, classification, function-calling agents, code completion, on-device assistants — start with Phi-4-mini (3.8B) and only size up if quality demands it. It's tiny, fast, supports 128K context and function calling, and costs nothing to run. Reach for the full Phi-4 (14B) when you need stronger STEM reasoning, and Phi-4-reasoning when the task is genuinely math/logic-heavy.

# Smallest viable: 3.8B, ~3-4GB VRAM, function calling, 128K context
ollama run phi4-mini

# STEM-dense flagship: 14B, ~8-10GB at Q4
ollama run phi4

For privacy-sensitive and regulated workloads

This is Phi-4's sweet spot. MIT license plus fully local inference means you can deploy it inside an air-gapped environment, fine-tune it on proprietary data without that data ever leaving your control, and ship it in a product without per-seat API costs. For document processing, internal copilots, and compliance-bound automation, the open-weight model is often the only viable option — and Phi-4 is among the most capable in its size class.

For multimodal and reasoning work on a budget

Phi-4-reasoning-plus gives you local AIME-2025-81% math reasoning on a single consumer GPU. Phi-4-reasoning-vision-15B handles diagrams, charts, and UI grounding well — useful for document AI, screen automation, and STEM-tutoring use cases — as long as you accept it's a specialist and not a general vision model. For broad multimodal understanding, Gemma 3 or a Qwen VL model is the better pick.

The routing discipline

The mental model: Phi-4 is the local-first default for well-scoped, high-volume, or sensitive work; a frontier API is the escalation path for the genuinely hard. Don't run a 14B model on problems that need a frontier model, and don't pay frontier API prices for work a 3.8B model nails. Match the model to the cost-of-error, not to the leaderboard.

FAQ

Is there a Phi-5 yet?

No. As of June 2026 there is no generally available Phi-5 with an official Microsoft model card. The current generation is the Phi-4 family, with the newest member being Phi-4-reasoning-vision-15B (March 4, 2026). Any "Phi-5" guides you find are speculative or based on pre-release rumor — verify against an official model card before trusting specs.

How much VRAM do I need to run Phi-4?

For the 14B model at Q4_K_M (the recommended quant), roughly 8-10GB of VRAM — it fits comfortably on a 10-12GB GPU. A Q3 quant squeezes into 8GB with headroom. The 3.8B Phi-4-mini needs only ~3-4GB and runs on most laptop GPUs, integrated graphics with enough shared memory, or CPU-only at reduced speed. Q4_K_M retains roughly 95% of full-precision quality on reasoning tasks.

What's the license, and can I use it commercially?

The Phi-4 family is open-weight under the MIT license — one of the most permissive available. You can use it commercially, fine-tune it, redistribute it, and ship it in products with no royalty. This is more permissive than Gemma's custom terms or the Llama license.

How does Phi-4 beat models larger than itself?

Microsoft trains Phi on "textbook-quality" curated and synthetic data rather than maximizing scraped tokens — the thesis is that data quality beats parameter count. The result: base Phi-4 (14B) outscores GPT-4o (its synthetic-data teacher) on GPQA and MATH, and Phi-4-reasoning (14B) beats 70B distillations and o1-mini on AIME. The tradeoff is breadth — Phi models are STEM-dense specialists, not generalists with frontier-level coverage of everything.

Where can I run Phi-4 models?

Ollama (ollama run phi4), LM Studio, llama.cpp, ONNX Runtime, Microsoft Foundry Local, Azure AI Foundry (managed endpoints), NVIDIA NIM (for Phi-4-mini), and raw weights on Hugging Face for fine-tuning. The same GGUF file is portable across the local-inference ecosystem.

Which benchmark numbers are verified vs vendor-claimed?

Base Phi-4's MMLU 84.8 / GPQA 56.1 / MATH 80.4 / HumanEval 82.6 come from Microsoft's technical report via OpenAI's simple-evals and are corroborated on llm-stats and the HF card — well-supported. The reasoning scores (AIME, GPQA Diamond) and all the vision scores (MathVista, MMMU, etc.) are from Microsoft's own technical reports and should be treated as vendor-claimed until broadly reproduced by third parties.

Analysis by Frank — former Oracle AI architect who helped build Oracle's AI Center of Excellence, now building agentic systems independently and making music with AI. Published June 5, 2026 with specs validated against Microsoft's Phi-4 and Phi-4-reasoning technical reports, the Hugging Face model cards, llm-stats, and community self-hosting guides. Vendor-claimed figures are marked as such, and no number here was invented.