MultimodalVery lowEstimatedReleased 2025-04-15

Phi-4 Multimodal Environmental Impact

First multimodal Phi — speech + vision + text, #1 on OpenASR

ArchitectureMultimodal Transformer (speech + vision + text)Parameters5.6BContext128,000 tokensProviderMicrosoft

The figures below are for one quick question. A longer chat with follow-ups uses about 0.17 Wh, and an hour or two of agent work about 4.9 Wh.

One query, in context

Energy, carbon and water for a single use. Quick single question (one exchange)

Energy0.05Wh per query

16% of a Google search.

likely 0.02 Wh – 0.10 Wh
Carbon0.02g per query

17% of a Google search.

likely 0.01 g – 0.04 g
Water0.18mL per query

About 3.6× a drop of water and 4% of a teaspoon of water.

likely 0.09 mL – 0.36 mL
  • Where it runsSelf-hosted / Azure
  • Grid carbon intensity458 g CO₂/kWh · 32% renewable
  • Compared with a Google searchabout the same as a search (0.3 Wh)
  • Water~5,556 queries to fill 1 litre

Where Phi-4 Multimodal sits

#16 of 199 by energy per query. Lighter than 92% of all models. Lit dots are the same kind of model; click any dot to open it.

Google searchKettle boilLLaMA 3.2 1B — 0.03 WhGemini 1.5 Pro — 0.45 WhGPT-4.1 Nano — 0.03 WhGPT-4o — 0.43 WhClaude 3.7 Sonnet — 0.21 WhGPT-4.5 — 1.0 Who3 (Reasoning) — 2.3 WhDeepSeek-R1 — 1.2 WhGrok 2 — 0.50 WhDALL-E 3 — 3.0 WhMidjourney — 7.5 WhStable Diffusion XL — 2.9 WhSora — 500 WhGitHub Copilot — 1.5 WhSuno — 75 WhGPT-4o Mini — 0.04 WhDeepSeek-V3 — 0.24 WhGemini 2.0 Flash — 0.08 WhClaude Opus 4 — 0.39 WhClaude Sonnet 4 — 0.21 WhClaude 3.5 Haiku — 0.06 WhMistral Large 2 — 1.8 WhLLaMA 3.1 70B — 0.38 WhPerplexity AI — 0.50 WhImagen 3 — 2.5 WhFlux.1 Dev — 3.8 WhVeo 2 — 350 WhGemini 2.5 Pro — 0.40 WhGemini 2.5 Flash — 0.24 WhGPT-4.1 — 0.37 WhGPT-4.1 Mini — 0.03 WhLLaMA 4 Scout — 0.11 WhLLaMA 4 Maverick — 0.11 WhQwen 3 235B — 0.16 WhGrok 3 — 0.40 WhGemma 3 27B — 0.16 WhPhi-4 — 0.09 WhCommand R+ — 0.56 WhGPT-4o Image Generation — 4.5 WhAdobe Firefly — 3.2 WhIdeogram 2.0 — 3.5 WhRunway Gen-3 Alpha — 200 WhKling — 250 WhPika — 180 WhLuma Dream Machine — 160 WhUdio — 70 Who1 — 2.3 Who1-mini — 0.20 Who3-mini — 0.20 Who4-mini — 0.20 WhGPT-4 Turbo — 0.50 WhGemini 1.5 Flash — 0.10 WhGemini Nano — 0.01 WhClaude 3 Opus — 0.56 WhClaude 3 Haiku — 0.06 WhLLaMA 3.1 8B — 0.06 WhLLaMA 3.1 405B — 2.1 WhLLaMA 3.3 70B — 0.38 WhLLaMA 3.2 11B Vision — 0.08 WhLLaMA 3.2 90B Vision — 0.49 WhMistral 7B — 0.06 WhMixtral 8x7B — 0.09 WhMixtral 8x22B — 0.22 WhMistral Small 3 — 0.14 WhCodestral — 0.80 WhAmazon Nova Pro — 0.30 WhAmazon Nova Lite — 0.06 WhAmazon Nova Micro — 0.03 WhERNIE Bot 4.0 — 0.30 WhDoubao Pro — 0.30 WhGLM-4 — 0.19 WhKimi — 0.25 WhJamba 1.5 Large — 1.0 WhStable Diffusion 3.5 — 3.2 WhLeonardo AI — 3.5 WhHaiLuo AI — 220 WhStable Video Diffusion — 150 WhCursor — 1.2 WhAmazon Q Developer — 1.0 WhClaude Code — 2.0 WhWindsurf (Codeium) — 0.80 WhReplit Agent — 1.5 WhWhisper — 0.50 WhOpenAI TTS — 0.30 WhElevenLabs — 1.5 WhNotebookLM Audio Overview — 15 WhClaude Sonnet 4.5 — 0.34 WhClaude Opus 4.5 — 0.42 WhClaude Opus 4.6 — 0.40 WhClaude Sonnet 4.6 — 0.20 WhGPT-5 — 0.44 WhGPT-5 Mini — 0.05 WhGPT-5.2 — 0.61 WhGPT-5.2 Codex — 0.61 WhGPT-5.4 — 0.60 WhGPT-5.4 Mini — 0.03 Who3-pro — 6.0 WhGPT Image 1.5 — 3.8 WhSora 2 — 600 WhGemini 3 Pro — 0.40 WhGemini 3 Flash — 0.15 WhGemini 3.1 Pro — 0.42 WhGemma 3 4B — 0.04 WhGemma 3 12B — 0.08 WhImagen 4 — 2.8 WhVeo 3 — 400 WhVeo 3.1 — 2.4 kWhGrok 4 — 0.60 WhGrok 4.1 — 0.50 WhDeepSeek-R1-0528 — 1.2 WhDeepSeek-V3.1 — 0.24 WhQwen3.5 (397B-A17B) — 0.13 WhQwen3.5 35B-A3B — 0.04 WhMistral Large 3 — 1.5 WhDevstral 2 — 0.60 WhVoxtral TTS — 0.40 WhKimi K2.5 — 0.30 WhGLM-5 — 0.22 WhDoubao 2.0 — 0.30 WhCommand A — 0.60 WhPhi-4 Mini — 0.04 WhPhi-4 Reasoning — 0.27 WhAmazon Nova 2 Pro — 0.35 WhMidjourney V7 — 8.0 WhFlux 2 Pro — 5.0 WhRunway Gen-4.5 — 250 WhKling 3.0 — 280 WhDevin — 3.0 WhElevenLabs v3 — 1.8 WhSuno v5.5 — 80 WhClaude Haiku 4.5 — 0.02 WhGPT-5.3 Codex — 0.45 WhGPT-5.4 Nano — 0.03 WhGPT-OSS 120B — 0.03 WhGPT-OSS 20B — 0.02 WhGemini 3.1 Flash Lite — 0.06 WhLyria 3 Pro — 50 WhDeepSeek-V3.2 — 0.25 WhGrok 4.20 — 0.50 WhQwen3-Coder 480B — 0.20 WhQwen 3.5 Omni — 0.20 WhNemotron 3 Super — 0.45 WhDoubao Seed 2.0 Pro — 0.35 WhSeedance 2.0 — 280 WhGLM-5.1 — 0.22 WhERNIE 5.0 — 0.35 WhMistral Small 4 — 0.40 WhMidjourney V8 Alpha — 5.0 WhRecraft V4 — 3.0 WhLuma Ray 3 — 200 WhPika 2.5 — 210 WhGPT-5.5 — 0.51 WhGPT-5.5 Instant — 0.06 WhGPT-5.3 Instant — 0.05 WhGPT Image 2 — 5.0 WhClaude Opus 4.7 — 0.44 WhClaude Opus 4.8 — 0.53 WhClaude Fable 5 — 0.79 WhClaude Sonnet 5 — 0.24 WhDeepSeek-V4-Pro — 0.24 WhDeepSeek-V4-Flash — 0.10 WhNano Banana 2 — 1.5 WhQwen3.6 27B — 0.16 WhQwen3.7-Max — 0.45 WhSora 2 Pro — 2.1 kWhClaude Opus 5 — 0.46 WhClaude Fable 5.1 — 0.65 WhClaude Opus 5.5 — 0.45 WhClaude Sonnet 5.5 — 0.24 WhClaude Haiku 5.5 — 0.02 WhClaude Fable 5.2 — 0.60 WhGPT-5.6 Sol — 0.39 WhGPT-5.6 Terra — 0.16 WhGPT-5.6 Luna — 0.05 WhGPT-6 Astra — 0.90 WhGPT-6 Sol — 0.42 WhGPT-6 Luna — 0.06 WhGPT-6 Instant — 0.07 WhGPT-6.1 Sol — 0.42 WhGPT-6.1 Terra — 0.17 WhGemini 3.5 Flash — 0.24 WhGemini 3.5 Flash-Lite — 0.06 WhGemini 3.6 Flash — 0.21 WhGemini 3.7 Flash — 0.14 WhGemini 3.8 Flash — 0.15 WhGemini 4 Argon — 0.45 WhNano Banana 2.1 — 1.5 WhGrok 4.5 — 0.35 WhGrok 4.7 — 0.40 WhMuse Spark — 0.40 WhMuse Spark Open — 0.22 WhKimi K3 — 0.65 WhGLM-5.3 — 0.22 WhQwen3.8-Max — 0.55 WhDeepSeek-V4.1-Flash — 0.10 WhMistral Large 4 — 1.4 WhKling 4.0 — 300 WhSeedance 2.5 — 290 Wh0.1 Wh1 Wh10 Wh100 Wh1 kWhPhi-4 Multimodal

Want something lighter?

Phi-4 Multimodal is already among the lighter well-known options for this kind of work, so there isn’t a clearly lighter alternative to suggest.

What a month of use adds up to

Pick how much you use it. Based on one quick use at a time, five days a week.

Energy
5.3 Wh
Carbon
2.2 g
Water
19 mL

A month of 5 uses a day, 5 days a week. The energy is 44% of a phone charge.

Work out your own →
Energy
10.6 Wh
Carbon
4.3 g
Water
39 mL

A month of 10 uses a day, 5 days a week. The energy is 88% of a phone charge.

Work out your own →
Energy
42.4 Wh
Carbon
17.3 g
Water
156 mL

A month of 40 uses a day, 5 days a week. The energy is about 3.5× a phone charge and 39% of a kettle boil.

Work out your own →

How does Phi-4 Multimodal compare?

Next to similar models. Bars use a log scale, so each step is a big jump.

Phi-4 Multimodal0.05 Wh

Detailed breakdown

Energy

We estimate Phi-4 Multimodal at ~0.049 Wh for a quick single question. The open-weight models CLEER has benchmarked directly follow a clear pattern: roughly 0.02 Wh plus 0.005 Wh per billion parameters active on each token (our fit across 11 models from 3B to 49B active). Phi-4 Multimodal activates ~5.6B (it is a dense model, so every parameter is active), which gives ~0.049 Wh when served from a batched data-centre deployment. Image and audio inputs add encoding work on top of this text-only figure. Run locally, the figure depends entirely on your hardware. This is our extrapolation, not a CLEER figure or a measurement. It replaces our earlier ~0.12 Wh estimate, which was built on 2025 benchmark figures for older hardware.

Power source & carbon

Open-source (MIT). Runs on mobile devices and edge hardware.

Water use

At ~0.18 mL per quick question, derived from the energy estimate at the ~3.8 L/kWh water factor we apply to this model's infrastructure. Water scales directly with energy, so longer chats, reasoning and agentic sessions use proportionally more.

About Phi-4 Multimodal

Phi-4 Multimodal is an open-source multimodal model from Microsoft, released in April 15, 2025, that runs well below the category average for energy consumption at 0.05 Wh per query. Because its weights are publicly available, it can be self-hosted on any infrastructure — meaning its carbon footprint depends entirely on where and how you choose to run it. At 5.6B parameters, it first multimodal phi — speech + vision + text, #1 on openasr.

These figures are estimates derived from hardware specifications and API benchmarks — Microsoft has not published official energy data for Phi-4 Multimodal. Actual consumption may vary significantly depending on batching, quantisation, and infrastructure optimisations that we cannot observe from outside.

Phi-4 Multimodal in context

0.4 kWhper year

Your yearly Phi-4 Multimodal footprint

At 25 queries per day, your annual Phi-4 Multimodal usage consumes 0.4 kWh — less than charging your phone for a year. That produces 0.2 kg of CO₂.

Key insights

  • Uses less than a third of the average energy for multimodal models
  • Open-source weights — can be self-hosted on infrastructure you control

The Microsoft Phi family

One chart per tier, on a shared time axis. Hover a dot for details; click to open that version.

Phi-4: 0.09 Wh · Dec 2024Phi-40.09 WhPhi-4 Mini: 0.04 Wh · Feb 2025Phi-4 Mini0.04 WhPhi-4 Multimodal: 0.05 Wh · Apr 2025Phi-4 Multimodal0.05 WhPhi-4 Reasoning: 0.27 Wh · May 2025Phi-4 Reasoning0.27 WhDec 2024May 2025
Phi-4 to Phi-4 Reasoning: 190% higher energy per query across 4 versions.

What does your Phi-4 Multimodal usage cost the planet?

Use the calculator to estimate your personal footprint based on how often you use Phi-4 Multimodal.

Calculate my compute

Frequently asked questions

How much energy does Phi-4 Multimodal use per query?

Each Phi-4 Multimodal query consumes approximately 0.05 Wh of energy. This is about the same as a traditional Google search (~0.3 Wh).

What is Phi-4 Multimodal's carbon footprint?

Based on the carbon intensity of Self-hosted / Azure, each query produces approximately 0.02 g of CO2. The grid in this region has a carbon intensity of 458 g CO2/kWh with 32% renewable energy.

How much water does Phi-4 Multimodal use?

Each query consumes approximately 0.18 mL of water, primarily used for cooling the data centers that process the request.

How does Phi-4 Multimodal compare to a Google search?

A Phi-4 Multimodal query uses about the same as a Google search in terms of energy. A Google search uses approximately 0.3 Wh, while Phi-4 Multimodal uses 0.05 Wh.

Technical details

Architecture
Multimodal Transformer (speech + vision + text)
Parameters
5.6B
Context window
128,000 tokens
Release date
2025-04-15
Open source
Yes
Training data cutoff
2025-03

Sources