About Phi-4
Phi-4 is Microsoft Research's argument that data quality beats parameter count: a compact dense model trained heavily on curated and synthetic textbook-style data, released under MIT, with reasoning and mathematics performance Microsoft highlighted as far above its weight class. It is text-only and deliberately small, the product of a research programme about how capable a carefully-fed small model can be.
Its niche is reasoning-dense work at the small end — maths tutoring, structured problem solving, education and research settings — rather than general assistant breadth.
Phi-4 is Microsoft's 14.7B-parameter model released in December 2024, with a 16K-token context window. It uses a classic dense-attention design whose KV cache grows linearly with context: its KV cache is about 1.7 GB at an 8K context, 3.4 GB at 16K, and 3.4 GB at the full 16K window (FP16 cache).
Memory is the easy part: quantised weights sit comfortably on mainstream gaming cards and recent laptops, with headroom to spare on anything mid-range or better. The constraint that actually shapes usage is its short context window — a fraction of what contemporaries offer — so long-document work and sprawling agent sessions are out of scope regardless of VRAM, and the calculator above will show context options capping out early. Within its lane it is one of the highest quality-per-gigabyte options available, and the MIT licence plus modest footprint make it a favourite for embedded and research deployments.
For most people Q4_K_M is the sweet spot — the most popular quality/size trade-off — while Q8 is near-lossless if you have the memory. Totals above include the KV cache and a realistic framework overhead, so they are what you should expect to see in practice rather than just the download size. Weight sizes are calibrated against real GGUF files — see the methodology.
Frequently asked questions
How much VRAM does Phi-4 need?
At Q4_K_M with an 8K context, Phi-4 needs about 13 GB (weights 9 GB + KV cache + overhead). The smallest common hardware that fits is a RTX 5060 Ti 16GB.
Can an RTX 4090 (24GB) run Phi-4?
Yes. An RTX 4090's 24 GB runs Phi-4 at Q8_0 (about 20 GB at 8K context).
Can a Mac run Phi-4?
Yes — Apple Silicon with 24 GB of unified memory or more (macOS lets the GPU use ~75% of it, ~18 GB) runs Phi-4 at Q4_K_M.
How much VRAM does Phi-4 need at its full 16K context?
The KV cache grows with context: about 1.7 GB at 8K tokens versus 3.4 GB at the full 16K window (FP16 cache). Add that to the weights (9 GB at Q4_K_M) plus overhead — long contexts can cost more than a whole quantisation step.
Related
VRAM calculator for any model ·Token counter
Similar-size local models: Gemma 4 12B (12B) · GPT-OSS 20B (21B) · Gemma 4 26B-A4B (26B) · DeepSeek R1 Distill Qwen3 8B (8.2B) · Llama 3.1 8B (8.03B)
Last updated 2026-09-03. Architecture figures from the model's published config.json; see the methodology.